[ { "path": ".agents/skills/fix-art-issues/SKILL.md", "content": "---\nname: fix-art-issues\ndescription: >\n Fix a GitHub issue on OpenPipe/ART and open a PR.\n Use when the user asks to fix, solve, or work on an ART issue,\n or references a GitHub issue URL containing \"OpenPipe/ART\".\n Triggers: \"fix ART issue\", \"solve this issue\" with an OpenPipe/ART URL,\n \"work on ART #N\".\n---\n\n# Fix ART Issue\n\nFix a GitHub issue on `OpenPipe/ART` and open a PR.\n\n- **Repo**: `OpenPipe/ART`\n- **Base branch**: `main`\n\nAssumes the workspace is already set up with the correct branch checked out and `.env` in place (handled by the system-level `fix-art-workspace` skill).\n\n## Workflow\n\n### 1. Read the Issue\n```\ngh issue view --repo OpenPipe/ART --json title,body,labels,assignees,comments\n```\n\n### 2. Explore, Plan, Implement\n- Use the Explore agent to understand relevant code before making changes.\n- Plan clearly, implement with minimal focused changes. No over-engineering.\n\n### 3. Commit and Push\n- Commit with a message that includes `Closes #`.\n- Push the feature branch. If HTTPS push fails due to SAML SSO, set SSH remote: `git remote set-url origin git@github.com:OpenPipe/ART.git`\n\n### 4. Open a Draft PR\n- `gh pr create --base main --draft`.\n- PR body: `## Summary`, `Closes #`, `## Changes`, `## Test plan`.\n\n### 5. Testing\n- **No test artifacts in the final PR**: debug prints, test scripts, and temporary changes must NOT be committed.\n- Update the PR's test plan section with detailed results.\n- When testing passes, mark the PR as ready: `gh pr ready`.\n\n## Reference\n\nRead `CONTRIBUTING.md` at the repo root for guidance on code quality checks (prek), CI cache refresh, and the release process.\n\n## Dependency Management Tips\n\n- **Pin versions strictly** (`==`) for critical deps like `transformers`, `trl`, `unsloth`, `unsloth-zoo`, `vllm` to avoid surprise breakage from new releases.\n- **Don't loosen pins without reason**: if a dep was `==X.Y.Z`, keep it pinned unless there's a specific reason to change. Don't use `>=` just because it seems more flexible.\n- **`uv run` fails on macOS** for backend deps (apex/torch need CUDA). This is expected — use `uvx ruff` for linting locally, test on GPU cluster.\n\n## Deploying a GPU Cluster\n\nName the SkyPilot cluster after the branch name without the `fix/` prefix, replacing `/` with `-` (SkyPilot doesn't allow slashes). For example, if the branch is `fix/short-description`:\n```\nuv run sky launch -c short-description skypilot-config.yaml -y\n```\n\nTo connect: `ssh short-description`\n\nTo tear down when done: `uv run sky down short-description`\n\n## GPU Cluster Testing Tips\n\n- **Kill stale GPU processes** before re-running tests: `nvidia-smi --query-compute-apps=pid --format=csv,noheader | xargs -r kill -9`. Previous failed runs leave processes holding GPU memory.\n- **Set `gpu_memory_utilization`** in test scripts (e.g. `0.7`) — the default `0.9` is too high when Unsloth's training model is also loaded on the same GPU.\n- **Redirect test output to a log file**: `nohup python test.py > /tmp/output.log 2>&1 &` then `tail -f /tmp/output.log`. SSH background tasks lose output when connection drops.\n- **Git on cluster**: SSH keys may not be configured. Use HTTPS with token: `git remote set-url origin https://${GITHUB_TOKEN}@github.com/OpenPipe/ART.git`\n- **Tear down clusters** when done: `sky down -y`\n\n$ARGUMENTS\n" }, { "path": ".agents/skills/train-rl/SKILL.md", "content": "---\nname: train-rl\ndescription: RL training reference for the ART framework. Use when the user asks to create, write, or help with an RL training script, reinforcement learning, GRPO, reward functions, RULER scoring, rollout functions, or anything related to RL fine-tuning.\n---\n\n# RL Training Wizard\n\nYou are guiding the user through setting up Reinforcement Learning (RL) training for a language model using the ART framework. Act as an interactive wizard: ask questions, validate inputs, and generate a complete runnable script.\n\n**Important**: Ask ONE question at a time. Wait for the user's response before asking the next question. Never bundle multiple questions into a single message.\n\n**Adaptability note**: Some steps reference tools like AskUserQuestion, Glob, or Bash. If you don't have access to these tools, simply ask the user the same questions as plain text and skip any steps that require running code (e.g., file search, dataset validation, hyperparameter computation). Do NOT fabricate results — never pretend you ran a tool or searched for files when you didn't.\n\n## Step 1: Single-turn or Multi-turn\n\nAsk the user using AskUserQuestion:\n\n1. **Single-turn** — The agent responds to a prompt once. A reward is assigned based on that single response (e.g., solve a math problem, classify text, answer a question).\n2. **Multi-turn** — The agent interacts over multiple turns with an environment, tools, or a game. A reward is assigned at the end of all turns (e.g., play a board game, use tools to complete a task, navigate a conversation).\n\n## Step 2: Describe the Task\n\nAsk the user to describe what the agent needs to do. Tell them you will help create a draft of the rollout function and environment, but they will likely need to edit it once the script is generated.\n\nGather:\n- **Task description** — What does the agent need to accomplish?\n- **Scenarios/inputs** — How are training inputs generated or provided? (e.g., a list of problems, a game generator, a dataset of tasks)\n- **System prompt** (optional) — Any system-level instructions for the agent\n\nFor **multi-turn** scenarios, also ask:\n- **Does the agent use tool calling?** — If yes, gather tool names, descriptions, parameter schemas (OpenAI function calling format), and how tool calls are executed (local function, API, MCP server, etc.)\n- **How does the environment work?** — What observations does the agent receive? What actions can it take? How does a turn work?\n- **When does an episode end?** — Win/loss conditions, turn limits, max tool calls, etc. (default max turns: 10)\n\nHelp the user flesh out incomplete descriptions. Offer to write helper functions (game logic, tool execution, scenario generators) as part of the final script.\n\n## Step 3: Reward Method\n\nAsk the user using AskUserQuestion:\n\n1. **Programmatic reward** — You have a ground truth or scoring function to compute the reward (e.g., check correctness against an answer, game win/loss, composite score)\n2. **RULER (LLM-as-judge)** — An LLM judge scores and compares the trajectories. No manual reward function needed. Requires an OpenAI API key (`OPENAI_API_KEY` env var).\n\nIf they choose **programmatic reward**, help them design a reward function. Common patterns:\n- **Binary**: 1 for correct, 0 for incorrect\n- **Accuracy**: fraction of correct sub-answers (0.0 to 1.0)\n- **Game outcome**: 1 for win, 0.5 for draw, 0 for loss, -1 for invalid move\n- **Scaled score**: logarithmic or normalized continuous score\n- **Composite**: weighted combination of multiple signals\n\nThe reward must be a float assigned to `trajectory.reward`. Additional signals can go in `trajectory.metrics` for W&B logging. **Important: `metrics` values must be numeric (`float`, `int`) or `bool` — strings are not allowed and will cause a Pydantic validation error.**\n\nIf they choose **RULER**, ask for:\n- **Judge model**: Recommend `openai/o4-mini` (default) or `openai/o3` for higher quality\n\n## Step 4: Gather Base Parameters\n\nDo NOT ask the user to review or confirm their answers after collecting them — just proceed to the next step.\n\n- **Base model**: Recommend ONLY these models:\n - `OpenPipe/Qwen3-14B-Instruct`\n - `Qwen/Qwen3-30B-A3B-Instruct-2507`\n - `meta-llama/Llama-3.1-8B-Instruct`\n- **Project name**: A name for this training project (default: `rl-project`)\n- **Run name**: A static, descriptive name (e.g., `math-solver-001`, `game-agent-001`). Ask the user for a meaningful name. Do NOT generate random names.\n\n## Step 5: Gather Hyperparameters\n\nPresent these defaults to the user, then ask using AskUserQuestion:\n- **Use defaults (Recommended)** — show all values in the description\n- **Customize** — adjust individual hyperparameters\n\nDefault values:\n- **Learning rate**: `1e-5`\n- **Number of training steps**: `50`\n- **Rollouts per group**: `8` (number of trajectories per scenario per step; more = better advantage estimation but slower). For RULER, default to `16`.\n- **Groups per step**: `1` (number of different scenarios per training step)\n\nIf they choose \"Customize\", ask which parameters to change.\n\n## Step 6: Generate the Training Script\n\nWrite a complete, runnable Python script by combining the appropriate **rollout pattern** (from Step 1/2) with the appropriate **reward method** (from Step 3) and the **training loop**.\n\nEvery script MUST:\n- Call `await backend.close()` at the end so the process doesn't hang\n- Print post-training info and usage examples (see shared block below)\n\n### Post-training block (append to ALL scripts before `backend.close()`):\n```python\n # --- Training complete ---\n step = await model.get_step()\n inference_name = model.get_inference_name()\n client = model.openai_client()\n\n print(\"\\n\" + \"=\" * 60)\n print(\"RL TRAINING COMPLETE\")\n print(\"=\" * 60)\n print(f\" Model: {inference_name}\")\n print(f\" Base model: \")\n print(f\" Training step: {step}\")\n print(f\" Inference URL: {client.base_url}\")\n print(\"=\" * 60)\n\n print(\"\\n--- Python usage (openai SDK) ---\\n\")\n print(f'''\\\nfrom openai import OpenAI\n\nclient = OpenAI(\n base_url=\"{client.base_url}\",\n api_key=\"not-needed\",\n)\n\nresponse = client.chat.completions.create(\n model=\"{inference_name}\",\n messages=[\n {{\"role\": \"user\", \"content\": \"Your prompt here\"}},\n ],\n)\nprint(response.choices[0].message.content)\n''')\n\n print(\"--- curl usage ---\\n\")\n print(f'''\\\ncurl {client.base_url}chat/completions \\\\\n -H \"Content-Type: application/json\" \\\\\n -d '{{\n \"model\": \"{inference_name}\",\n \"messages\": [\n {{\"role\": \"user\", \"content\": \"Your prompt here\"}}\n ]\n }}'\n''')\n\n await backend.close()\n```\n\n### Rollout pattern: Single-turn\n```python\nasync def rollout(model: art.Model, scenario: dict) -> art.Trajectory:\n client = AsyncOpenAI(\n base_url=model.inference_base_url,\n api_key=model.inference_api_key,\n )\n messages = [\n # {\"role\": \"system\", \"content\": \"\"},\n {\"role\": \"user\", \"content\": scenario[\"prompt\"]},\n ]\n response = await client.chat.completions.create(\n model=model.get_inference_name(),\n messages=messages,\n temperature=0.7,\n )\n choice = response.choices[0]\n\n # --- Compute reward (if programmatic) ---\n reward = # e.g., 1.0 if correct else 0.0\n\n return art.Trajectory(\n messages_and_choices=[*messages, choice],\n reward=reward,\n metrics={\"acc\": reward},\n )\n```\n\n### Rollout pattern: Multi-turn (environment/game loop)\n```python\nasync def rollout(model: art.Model, scenario) -> art.Trajectory:\n client = AsyncOpenAI(\n base_url=model.inference_base_url,\n api_key=model.inference_api_key,\n )\n game = create_game()\n traj = art.Trajectory(\n messages_and_choices=[\n {\"role\": \"system\", \"content\": \"\"},\n ],\n reward=0.0,\n )\n\n while not is_finished(game):\n traj.messages_and_choices.append(\n {\"role\": \"user\", \"content\": render_observation(game)}\n )\n response = await client.chat.completions.create(\n model=model.get_inference_name(),\n messages=traj.messages(),\n temperature=0.7,\n max_completion_tokens=256,\n )\n choice = response.choices[0]\n traj.messages_and_choices.append(choice)\n try:\n apply_action(game, choice.message.content)\n except ValueError:\n traj.reward = -1.0\n return traj\n\n traj.reward = compute_reward(game)\n return traj\n```\n\n### Rollout pattern: Multi-turn with tool calling\n```python\nasync def rollout(model: art.Model, scenario: dict) -> art.Trajectory:\n client = AsyncOpenAI(\n base_url=model.inference_base_url,\n api_key=model.inference_api_key,\n )\n MAX_TURNS = \n traj = art.Trajectory(\n messages_and_choices=[\n # {\"role\": \"system\", \"content\": \"\"},\n {\"role\": \"user\", \"content\": scenario[\"task\"]},\n ],\n tools=tools,\n reward=0.0,\n )\n\n for turn in range(MAX_TURNS):\n response = await client.chat.completions.create(\n model=model.get_inference_name(),\n messages=traj.messages(),\n tools=tools,\n temperature=0.7,\n )\n choice = response.choices[0]\n traj.messages_and_choices.append(choice)\n\n if choice.message.tool_calls:\n for tc in choice.message.tool_calls:\n args = json.loads(tc.function.arguments)\n result = execute_tool(tc.function.name, args)\n traj.messages_and_choices.append({\n \"role\": \"tool\",\n \"tool_call_id\": tc.id,\n \"content\": str(result),\n })\n else:\n break # Agent finished (no more tool calls)\n\n # --- Compute reward (if programmatic) ---\n traj.reward = \n return traj\n```\n\n### Reward method: RULER addition\nWhen using RULER, the rollout function should set `reward=0.0` (RULER fills it in). Add this scoring block inside the training loop, after `gather_trajectory_groups` and before `model.train`:\n\n```python\nfrom art.rewards import ruler_score_group\n\n # Score with RULER (LLM judge assigns relative rewards 0-1)\n judged_groups = []\n for group in finished_groups:\n judged = await ruler_score_group(\n group,\n judge_model=JUDGE_MODEL,\n debug=True,\n )\n judged_groups.append(judged)\n finished_groups = judged_groups\n```\n\n### Training loop (shared by all patterns):\n```python\n\"\"\"RL training script generated by /train-rl wizard.\"\"\"\nimport asyncio\nimport json\nfrom openai import AsyncOpenAI\nimport art\nfrom art.local import LocalBackend\n\n# --- Scenarios ---\nscenarios = [\n # Define or load your training scenarios here.\n]\n\n# --- Rollout function ---\n# (insert the appropriate rollout pattern here)\n\n# --- Training loop ---\nasync def main():\n backend = LocalBackend()\n model = art.TrainableModel(\n name=\"\",\n project=\"\",\n base_model=\"\",\n _internal_config=art.dev.InternalModelConfig(\n engine_args={\"gpu_memory_utilization\": 0.7},\n ),\n )\n await model.register(backend)\n\n NUM_STEPS = \n ROLLOUTS_PER_GROUP = \n GROUPS_PER_STEP = \n\n for step in range(await model.get_step(), NUM_STEPS):\n groups = [\n art.TrajectoryGroup(\n rollout(model, scenarios[\n (step * GROUPS_PER_STEP + i) % len(scenarios)\n ])\n for _ in range(ROLLOUTS_PER_GROUP)\n )\n for i in range(GROUPS_PER_STEP)\n ]\n finished_groups = await art.gather_trajectory_groups(\n groups, pbar_desc=f\"step {step}\"\n )\n\n # (insert RULER scoring block here if using LLM-as-judge)\n\n avg_reward = sum(\n t.reward for g in finished_groups for t in g.trajectories\n ) / max(1, sum(len(g.trajectories) for g in finished_groups))\n print(f\"Step {step}: avg_reward={avg_reward:.3f}\")\n\n await model.delete_checkpoints()\n await model.train(\n finished_groups,\n config=art.TrainConfig(learning_rate=),\n )\n\n # ... post-training block + backend.close() ...\n\nif __name__ == \"__main__\":\n asyncio.run(main())\n```\n\n### Alternative loop: Dataset-driven with iterate_dataset\nWhen the user has a fixed list of training scenarios and wants epoch-based iteration, use `iterate_dataset` instead of the manual step loop. This can be combined with any rollout pattern and reward method.\n\n```python\nfrom art.utils import iterate_dataset\n\n # Replace the manual for-loop with:\n training_iterator = iterate_dataset(\n scenarios,\n groups_per_step=,\n num_epochs=,\n initial_step=await model.get_step(),\n )\n\n for batch in training_iterator:\n groups = [\n art.TrajectoryGroup(\n rollout(model, item) for _ in range(ROLLOUTS_PER_GROUP)\n )\n for item in batch.items\n ]\n finished_groups = await art.gather_trajectory_groups(\n groups, pbar_desc=f\"epoch {batch.epoch} step {batch.step}\"\n )\n\n # (insert RULER scoring block here if using LLM-as-judge)\n\n avg_reward = sum(\n t.reward for g in finished_groups for t in g.trajectories\n ) / max(1, sum(len(g.trajectories) for g in finished_groups))\n print(f\"Step {batch.step} (epoch {batch.epoch}): avg_reward={avg_reward:.3f}\")\n\n await model.delete_checkpoints()\n await model.train(\n finished_groups,\n config=art.TrainConfig(learning_rate=),\n )\n```\n\n## Step 7: Write and Offer to Run\n\n1. Write the script to a file (suggest `rl_train.py`)\n2. Ask the user if they want to run it now with `uv run python `\n3. If yes, run it **directly using the Bash tool** (do NOT delegate to a Task subagent) so training logs stream live to the user. Use a **2-minute timeout**. If it times out, check progress and decide whether to continue.\n4. **GPU memory errors**: If training fails with OOM, lower `gpu_memory_utilization` in the existing `_internal_config` (e.g. from `0.7` to `0.5`).\n5. **Stale GPU memory**: If available GPU memory looks too small, previous training runs may still be occupying memory. Before retrying, run `nvidia-smi` to check, and if needed kill leftover processes with `kill ` to free memory.\n\n## Important Notes\n\n- LocalBackend requires a GPU.\n- RL uses **GRPO** (Group Relative Policy Optimization) under the hood. It needs multiple trajectories per scenario (a `TrajectoryGroup`) to compute relative advantages. More rollouts per group = better advantage estimation.\n- **RULER** eliminates the need for manual reward engineering by using an LLM judge to compare trajectories within a group. It requires an OpenAI API key (`OPENAI_API_KEY` env var).\n- The `@art.retry` decorator can wrap rollout functions to handle transient errors: `@art.retry(exceptions=(openai.LengthFinishReasonError,))`.\n- **Validation**: To log validation metrics without training, use `await model.log(val_groups)` or `await model.log(val_groups, split=\"val\")`.\n- **Resuming**: All patterns use `await model.get_step()` as the loop start, so training resumes from the last checkpoint automatically.\n" }, { "path": ".agents/skills/train-sft/SKILL.md", "content": "---\nname: train-sft\ndescription: SFT training reference for the ART framework. Use when the user asks to create, write, or help with an SFT training script, fine-tune a model, train from a JSONL dataset, do distillation, or anything related to supervised fine-tuning.\n---\n\n# SFT Training Wizard\n\nYou are guiding the user through setting up Supervised Fine-Tuning (SFT) for a language model using the ART framework. Act as an interactive wizard: ask questions, validate inputs, and generate a complete runnable script.\n\n**Important**: Ask ONE question at a time. Wait for the user's response before asking the next question. Never bundle multiple questions into a single message.\n\n**Adaptability note**: Some steps reference tools like AskUserQuestion, Glob, or Bash. If you don't have access to these tools, simply ask the user the same questions as plain text and skip any steps that require running code (e.g., file search, dataset validation, hyperparameter computation). Do NOT fabricate results — never pretend you ran a tool or searched for files when you didn't.\n\n## Step 1: Determine Training Scenario\n\nAsk the user ONE question at a time. Wait for their response before moving to the next question.\n\n**Training scenario:**\n1. **Train from a JSONL file** — They have a dataset file with chat-formatted examples\n2. **Distillation** — They want to train a smaller model using outputs from a larger teacher model\n\n## Step 2: Determine Backend\n\n**Backend:**\n1. **ServerlessBackend (Recommended)** — Train on remote managed GPUs. No local GPU needed, production-ready inference endpoint.\n2. **LocalBackend** — Train on your local GPU. Full control, fast iteration.\n\n## Step 3: Select and Validate Dataset (JSONL scenario)\n\n**IMPORTANT**: Do NOT assume a dataset. Do NOT make up or hallucinate file paths. Never pretend you searched for files if you didn't actually run a search tool.\n\nIf you have access to file system tools (Glob) and can actually execute them, search for `.jsonl` files using Glob (`**/*.jsonl`). Present real results as options. Always include \"Provide my own file path\" as the last option.\n\nOtherwise, ask the user: \"What is the path to your JSONL training file?\" — nothing more.\n\nOnce the user has provided a file path, validate it if you can run code using the script below. If you cannot run code, skip validation and move on.\n\n```python\nimport json, sys\nROLES = {\"system\", \"user\", \"assistant\", \"developer\", \"tool\", \"function\"}\nerrors = []\nfor i, line in enumerate(open(sys.argv[1]), 1):\n try:\n r = json.loads(line)\n msgs = r.get(\"input\", r).get(\"messages\", [])\n assert isinstance(msgs, list) and msgs, \"no messages\"\n for j, m in enumerate(msgs):\n assert m.get(\"role\") in ROLES, f\"messages[{j}]: invalid role {m.get('role')!r}\"\n assert m.get(\"content\") or m.get(\"function_call\") or m.get(\"tool_calls\"), f\"messages[{j}]: no content\"\n if \"input\" not in r:\n assert msgs[-1][\"role\"] == \"assistant\", \"last message must be from assistant\"\n tools = r.get(\"tools\")\n if tools is not None:\n assert isinstance(tools, list), \"tools must be a list\"\n except Exception as e:\n errors.append(f\" Line {i}: {e}\")\nprint(f\"{len(errors)} error(s):\\n\" + \"\\n\".join(errors) if errors else f\"Valid! {i} rows\")\nsys.exit(1 if errors else 0)\n```\n\nThe JSONL format supports these fields per row:\n- **`messages`** (required): List of chat messages\n- **`tools`** (optional): List of tool/function definitions for tool-call training\n- **`response_format`** (optional): Structured output schema (not used during training, but useful as metadata)\n\nReport the row count and validation result to the user. Do NOT read the whole dataset file. Do NOT name the dataset. If the format is wrong, help them fix it or convert their data.\n\n## Step 4: Gather Base Parameters\n\nDo NOT ask the user to review or confirm their answers after collecting them — just proceed to the next step.\n\n- **Base model**: Recommend ONLY these models:\n - `OpenPipe/Qwen3-14B-Instruct`\n - `Qwen/Qwen3-30B-A3B-Instruct-2507`\n - `meta-llama/Llama-3.1-8B-Instruct`\n- **Project name**: A name for this training project (default: `sft-project`)\n- **Run name**: A static, descriptive name (e.g., `agent-001`, `pii-redactor-001`, `math-tutor-001`). Ask the user for a meaningful name. Do NOT generate random names.\n\nFor **distillation** also ask:\n- **Teacher model**: The larger model to distill from (e.g., an OpenRouter model)\n- **Teacher API base URL and key**: If using a third-party provider\n- **Prompts**: What prompts to send to the teacher model\n\n## Step 5: Gather Hyperparameters\n\nThis step only applies if you can run code AND know the row count from validation. If you cannot run code, skip this step entirely — do NOT make up or guess hyperparameter values. The `train_sft_from_file` function has sensible built-in defaults.\n\nRun this Python snippet via Bash to compute defaults (replace `NUM_ROWS` with the actual row count). Do NOT show any formulas or calculation steps to the user — only show the final values.\n\n```python\nimport math, sys\nn = int(sys.argv[1])\nepochs = max(1, min(10, round(10000 / n)))\nbatch_size = 2\ntotal_steps = math.ceil(n * epochs / batch_size)\nsteps_per_epoch = math.ceil(n / batch_size)\nwarmup_steps = max(10, min(1000, round(steps_per_epoch * 0.05)))\nwarmup_ratio = round(warmup_steps / total_steps, 4)\nprint(f\"epochs={epochs} batch_size={batch_size} lr=2e-4 schedule=linear warmup_ratio={warmup_ratio}\")\n```\n\nPresent the output values to the user, then ask:\n- **Use defaults (Recommended)** — show all values in the description\n- **Customize** — adjust individual hyperparameters\n\nIf they choose \"Customize\", ask which parameters to change.\n\n### For distillation:\nUse the same defaults computation as JSONL (replace `NUM_ROWS` with the number of trajectories). `create_sft_dataset_iterator` handles the LR schedule automatically.\n\n## Step 6: Generate the Training Script\n\nWrite a complete, runnable Python script. Use the patterns below. Every script MUST:\n- Call `await backend.close()` at the end so the process doesn't hang\n- Print post-training info and usage examples (see shared block below)\n\n### Post-training block (append to ALL scripts before `backend.close()`):\n```python\n # --- Training complete ---\n step = await model.get_step()\n inference_name = model.get_inference_name()\n client = model.openai_client()\n\n print(\"\\n\" + \"=\" * 60)\n print(\"SFT TRAINING COMPLETE\")\n print(\"=\" * 60)\n print(f\" Model: {inference_name}\")\n print(f\" Base model: \")\n print(f\" Training step: {step}\")\n print(f\" Inference URL: {client.base_url}\")\n print(f\" W&B run: https://wandb.ai///runs/\")\n print(\"=\" * 60)\n\n print(\"\\n--- Python usage (openai SDK) ---\\n\")\n print(f'''\\\nfrom openai import OpenAI\n\nclient = OpenAI(\n base_url=\"{client.base_url}\",\n api_key=\"not-needed\",\n)\n\nresponse = client.chat.completions.create(\n model=\"{inference_name}\",\n messages=[\n {{\"role\": \"user\", \"content\": \"Your prompt here\"}},\n ],\n)\nprint(response.choices[0].message.content)\n''')\n\n print(\"--- curl usage ---\\n\")\n print(f'''\\\ncurl {client.base_url}chat/completions \\\\\n -H \"Content-Type: application/json\" \\\\\n -d '{{\n \"model\": \"{inference_name}\",\n \"messages\": [\n {{\"role\": \"user\", \"content\": \"Your prompt here\"}}\n ]\n }}'\n''')\n\n await backend.close()\n```\n\n### Backend setup\n\nUse the appropriate backend based on the user's choice:\n\n**LocalBackend:**\n```python\nfrom art.local import LocalBackend\n\nbackend = LocalBackend()\nmodel = art.TrainableModel(\n name=\"\",\n project=\"\",\n base_model=\"\",\n _internal_config=art.dev.InternalModelConfig(\n engine_args={\"gpu_memory_utilization\": 0.7},\n ),\n)\nawait model.register(backend)\n```\n\n**ServerlessBackend:**\n```python\nfrom art.serverless.backend import ServerlessBackend\n\nbackend = ServerlessBackend() # uses WANDB_API_KEY env var\nmodel = art.TrainableModel(\n name=\"\",\n project=\"\",\n base_model=\"\",\n)\nawait model.register(backend)\n```\n\nNote: `_internal_config` with `gpu_memory_utilization` is only used with LocalBackend. Do NOT include it for ServerlessBackend.\n\n### JSONL file training pattern:\n\nIf hyperparameters were computed in Step 5, pass them explicitly. If Step 5 was skipped, omit them — `train_sft_from_file` has sensible defaults.\n\n```python\n\"\"\"SFT training script generated by /train-sft wizard.\"\"\"\nimport asyncio\nimport art\n\nfrom art.utils.sft import train_sft_from_file\n\nasync def main():\n \n\n await train_sft_from_file(\n model=model,\n file_path=\"\",\n # Only include these if hyperparameters were computed:\n # epochs=,\n # batch_size=,\n # peak_lr=,\n # schedule_type=\"\",\n # warmup_ratio=,\n verbose=True,\n )\n\n # ... post-training block + backend.close() ...\n\nif __name__ == \"__main__\":\n asyncio.run(main())\n```\n\n### Distillation pattern:\n```python\n\"\"\"Distillation SFT script generated by /train-sft wizard.\"\"\"\nimport asyncio, os\nfrom dotenv import load_dotenv\nfrom openai import AsyncOpenAI\nimport art\n\nfrom art.utils.sft import create_sft_dataset_iterator\n\nload_dotenv()\n\nasync def main():\n teacher_client = AsyncOpenAI(\n api_key=os.environ[\"\"],\n base_url=\"\",\n )\n prompts = [\"\", \"\"]\n\n trajectories = []\n for prompt in prompts:\n completion = await teacher_client.chat.completions.create(\n model=\"\",\n messages=[{\"role\": \"user\", \"content\": prompt}],\n )\n trajectories.append(\n art.Trajectory(\n messages_and_choices=[\n {\"role\": \"user\", \"content\": prompt},\n {\"role\": \"assistant\", \"content\": completion.choices[0].message.content},\n ],\n tools=,\n )\n )\n\n \n\n for chunk in create_sft_dataset_iterator(\n trajectories,\n epochs=,\n batch_size=,\n peak_lr=,\n schedule_type=\"\",\n warmup_ratio=,\n ):\n await model.train_sft(chunk.trajectories, chunk.config, verbose=True)\n\n # ... post-training block + backend.close() ...\n\nif __name__ == \"__main__\":\n asyncio.run(main())\n```\n\n## Step 7: Write and Offer to Run\n\n1. Write the script to a file (suggest `sft_train.py`)\n2. Ask the user if they want to run it now with `uv run python `\n3. If yes, run it **directly using the Bash tool** (do NOT delegate to a Task subagent) so training logs stream live to the user. Use a **2-minute timeout**. If it times out, check progress and decide whether to continue.\n4. **LocalBackend only — GPU memory errors**: If training fails with OOM, lower `gpu_memory_utilization` in the existing `_internal_config` (e.g. from `0.7` to `0.5`).\n5. **LocalBackend only — Stale GPU memory**: If available GPU memory looks too small, previous training runs may still be occupying memory. Before retrying, run `nvidia-smi` to check, and if needed kill leftover processes with `kill ` to free memory.\n\n## Important Notes\n\n- LocalBackend requires a GPU.\n- ServerlessBackend requires a `WANDB_API_KEY` environment variable.\n" }, { "path": ".dockerignore", "content": "__pycache__/\n.art/\n# .env\n.venv/\ngrpo_trainer_lora_model/\nlogs/\nshared_cache.db\nstreaming-chat-completions/\nunsloth_compiled_cache/\nwandb/\ndocs/node_modules/\ndist/\nreplays/\ntrajectories/\n.DS_Store\n# .local/\n# .claude/\n.vscode/" }, { "path": ".github/workflows/create-draft-release.yml", "content": "name: Create Draft Release\n\non:\n workflow_dispatch:\n inputs:\n version_type:\n description: 'Version bump type'\n required: true\n default: 'patch'\n type: choice\n options:\n - patch\n - minor\n - major\n\npermissions:\n contents: write\n pull-requests: write\n\njobs:\n create-draft:\n runs-on: ubuntu-latest\n steps:\n - uses: actions/checkout@v4\n with:\n fetch-depth: 0\n\n - name: Set up Python\n uses: actions/setup-python@v5\n with:\n python-version: '3.11'\n\n - name: Install uv\n run: |\n curl -LsSf https://astral.sh/uv/install.sh | sh\n echo \"$HOME/.cargo/bin\" >> $GITHUB_PATH\n\n - name: Install dependencies\n run: |\n uv venv\n uv pip install -e .\n\n - name: Bump version\n id: bump\n run: |\n python scripts/bump_version.py ${{ github.event.inputs.version_type }}\n NEW_VERSION=$(python -c \"import tomllib; print(tomllib.load(open('pyproject.toml', 'rb'))['project']['version'])\")\n echo \"NEW_VERSION=$NEW_VERSION\" >> $GITHUB_OUTPUT\n\n - name: Generate release notes template\n id: release_notes\n env:\n GH_TOKEN: ${{ github.token }}\n run: |\n # Get the previous tag\n PREVIOUS_TAG=$(git describe --tags --abbrev=0 2>/dev/null || echo \"\")\n \n # Get merged PRs since last release\n if [ -z \"$PREVIOUS_TAG\" ]; then\n PRS=$(gh pr list --state merged --limit 100 --json number,title,url | jq -r '.[] | \"- \\(.title) ([#\\(.number)](\\(.url)))\"')\n else\n LAST_TAG_DATE=$(git log -1 --format=%aI $PREVIOUS_TAG)\n PRS=$(gh pr list --state merged --limit 100 --json number,title,url,mergedAt | jq -r --arg date \"$LAST_TAG_DATE\" '.[] | select(.mergedAt > $date) | \"- \\(.title) ([#\\(.number)](\\(.url)))\"')\n fi\n \n # Create release notes template\n cat << EOF > release_notes_template.md\n ## Release Highlights\n \n \n \n ## What's Changed\n \n $PRS\n \n **Full Changelog**: https://github.com/OpenPipe/ART/compare/$PREVIOUS_TAG...v${{ steps.bump.outputs.NEW_VERSION }}\n EOF\n\n - name: Create draft release\n env:\n GH_TOKEN: ${{ github.token }}\n run: |\n gh release create v${{ steps.bump.outputs.NEW_VERSION }} \\\n --title \"v${{ steps.bump.outputs.NEW_VERSION }}\" \\\n --notes-file release_notes_template.md \\\n --draft\n\n - name: Create PR with version bump\n env:\n GH_TOKEN: ${{ github.token }}\n run: |\n git config --local user.email \"action@github.com\"\n git config --local user.name \"GitHub Action\"\n git checkout -b release/v${{ steps.bump.outputs.NEW_VERSION }}\n git add pyproject.toml uv.lock\n git commit -m \"Bump version to ${{ steps.bump.outputs.NEW_VERSION }}\"\n git push origin release/v${{ steps.bump.outputs.NEW_VERSION }}\n \n gh pr create \\\n --title \"Release v${{ steps.bump.outputs.NEW_VERSION }}\" \\\n --body \"This PR bumps the version to ${{ steps.bump.outputs.NEW_VERSION }}. \n\n **Next steps:**\n 1. Review and edit the [draft release](https://github.com/OpenPipe/ART/releases) \n 2. Add release highlights and curate the changelog\n 3. Merge this PR to publish the release automatically\" \\\n --base main \\\n --head release/v${{ steps.bump.outputs.NEW_VERSION }}\n\n - name: Output instructions\n run: |\n echo \"::notice::Draft release created! Next steps:\"\n echo \"::notice::1. Go to https://github.com/OpenPipe/ART/releases and edit the draft\"\n echo \"::notice::2. Add release highlights and curate the auto-generated PR list\" \n echo \"::notice::3. Merge the PR to publish the release\"\n" }, { "path": ".github/workflows/package-install.yml", "content": "name: Package Install\n\non:\n pull_request:\n push:\n branches: [main]\n workflow_dispatch:\n\npermissions:\n contents: read\n\njobs:\n install-smoke-test:\n runs-on: ubuntu-latest\n\n steps:\n - uses: actions/checkout@v4\n\n - name: Set up Python\n uses: actions/setup-python@v5\n with:\n python-version: \"3.11\"\n\n - name: Install uv\n run: |\n curl -LsSf https://astral.sh/uv/install.sh | sh\n echo \"$HOME/.cargo/bin\" >> \"$GITHUB_PATH\"\n\n - name: Build wheel\n run: uv build --wheel --out-dir dist\n\n - name: Smoke test uv add + sync for backend extra\n run: |\n wheel_path=\"$(python - <<'PY'\n from pathlib import Path\n\n print(next(Path(\"dist\").glob(\"openpipe_art-*.whl\")).resolve())\n PY\n )\"\n\n project_dir=\"$(mktemp -d)\"\n cd \"$project_dir\"\n uv init --name art-install-smoke --python 3.11 --bare\n uv add \"openpipe-art[backend] @ file://${wheel_path}\"\n uv sync\n" }, { "path": ".github/workflows/prek.yml", "content": "name: Prek\n\non:\n pull_request:\n push:\n branches: [main]\n\npermissions:\n contents: write\n\nenv:\n CI_BASE_IMAGE: \"pytorch/pytorch:2.9.0-cuda12.8-cudnn9-devel\"\n CI_PYTHON_MM: \"3.11\"\n CI_UV_CACHE_RELEASE_TAG: \"prek-uv-cache\"\n CI_UV_CACHE_ASSET_PREFIX: \"prek-uv-cache\"\n UV_CACHE_DIR: \"/root/.cache/uv\"\n UV_LINK_MODE: \"copy\"\n TORCH_CUDA_ARCH_LIST: \"8.0\"\n\njobs:\n cache-status:\n runs-on: art-large-runner\n outputs:\n cache-hit: ${{ steps.check.outputs.cache-hit }}\n fingerprint: ${{ steps.fingerprint.outputs.fingerprint }}\n steps:\n - name: Checkout code\n uses: actions/checkout@v4\n\n - name: Compute expected uv cache fingerprint\n id: fingerprint\n run: |\n fp=\"$(python3 scripts/ci/compute_uv_fingerprint.py \\\n --pyproject pyproject.toml \\\n --uv-lock uv.lock \\\n --base-image \"${CI_BASE_IMAGE}\" \\\n --python-mm \"${CI_PYTHON_MM}\")\"\n echo \"fingerprint=${fp}\" >> \"${GITHUB_OUTPUT}\"\n echo \"Expected uv cache fingerprint: ${fp}\"\n\n - name: Check if uv cache exists\n id: check\n env:\n GITHUB_TOKEN: ${{ secrets.GITHUB_TOKEN }}\n run: |\n fingerprint=\"${{ steps.fingerprint.outputs.fingerprint }}\"\n part_prefix=\"${CI_UV_CACHE_ASSET_PREFIX}-${fingerprint}.tar.zst.part-\"\n release_api=\"https://api.github.com/repos/${GITHUB_REPOSITORY}/releases/tags/${CI_UV_CACHE_RELEASE_TAG}\"\n\n release_json=\"$(curl -fsSL \\\n -H \"Authorization: Bearer ${GITHUB_TOKEN}\" \\\n -H \"Accept: application/vnd.github+json\" \\\n \"${release_api}\" || true)\"\n\n if [ -z \"${release_json}\" ]; then\n echo \"Cache release '${CI_UV_CACHE_RELEASE_TAG}' not found.\"\n echo \"cache-hit=false\" >> \"${GITHUB_OUTPUT}\"\n exit 0\n fi\n\n hit=\"$(RELEASE_JSON=\"${release_json}\" PART_PREFIX=\"${part_prefix}\" python3 -c \"\n import json, os, re\n payload = json.loads(os.environ['RELEASE_JSON'])\n prefix = os.environ['PART_PREFIX']\n pattern = re.compile(r'^' + re.escape(prefix) + r'(\\d{3})$')\n parts = sorted(\n int(m.group(1))\n for a in payload.get('assets', [])\n for m in [pattern.match(a.get('name', ''))]\n if m and a.get('id') is not None\n )\n print('true' if parts and parts == list(range(len(parts))) else 'false')\n \")\"\n echo \"cache-hit=${hit}\" >> \"${GITHUB_OUTPUT}\"\n echo \"Cache hit: ${hit}\"\n\n build-cache:\n needs: cache-status\n if: needs.cache-status.outputs.cache-hit != 'true'\n runs-on: art-cache-builder\n container:\n image: pytorch/pytorch:2.9.0-cuda12.8-cudnn9-devel\n steps:\n - name: Install CI dependencies\n run: |\n apt-get update\n apt-get install -y --no-install-recommends ca-certificates curl git zstd\n rm -rf /var/lib/apt/lists/*\n curl -LsSf https://astral.sh/uv/install.sh | sh\n echo \"/root/.local/bin\" >> \"${GITHUB_PATH}\"\n\n - name: Install gh CLI\n env:\n GH_TOKEN: ${{ secrets.GITHUB_TOKEN }}\n run: |\n GH_DL_URL=\"$(curl -fsSL \\\n -H \"Authorization: Bearer ${GH_TOKEN}\" \\\n https://api.github.com/repos/cli/cli/releases/latest \\\n | python3 -c \"import json,sys;r=json.load(sys.stdin);print([a['browser_download_url'] for a in r['assets'] if a['name'].endswith('_linux_amd64.tar.gz')][0])\")\"\n curl -fsSL \"${GH_DL_URL}\" | tar xz --strip-components=1 -C /usr/local\n gh version\n\n - name: Checkout code\n uses: actions/checkout@v4\n\n - name: Mark workspace as a safe git directory\n run: |\n git config --global --add safe.directory \"${GITHUB_WORKSPACE}\"\n\n - name: Build and upload uv cache\n env:\n GH_TOKEN: ${{ secrets.GITHUB_TOKEN }}\n run: |\n bash scripts/ci/build_and_push_uv_cache.sh \\\n --base-image \"${CI_BASE_IMAGE}\" \\\n --python-mm \"${CI_PYTHON_MM}\"\n\n quality-checks:\n needs: [cache-status, build-cache]\n if: ${{ !failure() && !cancelled() }}\n runs-on: art-large-runner\n container:\n image: pytorch/pytorch:2.9.0-cuda12.8-cudnn9-devel\n steps:\n - name: Install CI dependencies\n run: |\n apt-get update\n apt-get install -y --no-install-recommends ca-certificates curl git zstd\n rm -rf /var/lib/apt/lists/*\n curl -LsSf https://astral.sh/uv/install.sh | sh\n echo \"/root/.local/bin\" >> \"${GITHUB_PATH}\"\n\n - name: Checkout code\n uses: actions/checkout@v4\n\n - name: Mark workspace as a safe git directory\n run: |\n git config --global --add safe.directory \"${GITHUB_WORKSPACE}\"\n\n - name: Restore prebuilt uv cache\n env:\n GITHUB_TOKEN: ${{ secrets.GITHUB_TOKEN }}\n run: |\n release_api=\"https://api.github.com/repos/${GITHUB_REPOSITORY}/releases/tags/${CI_UV_CACHE_RELEASE_TAG}\"\n fingerprint=\"${{ needs.cache-status.outputs.fingerprint }}\"\n part_prefix=\"${CI_UV_CACHE_ASSET_PREFIX}-${fingerprint}.tar.zst.part-\"\n\n release_json=\"$(curl -fsSL \\\n -H \"Authorization: Bearer ${GITHUB_TOKEN}\" \\\n -H \"Accept: application/vnd.github+json\" \\\n \"${release_api}\" || true)\"\n\n if [ -z \"${release_json}\" ]; then\n echo \"::error::Missing cache release '${CI_UV_CACHE_RELEASE_TAG}'.\"\n exit 1\n fi\n\n part_selection_file=\"/tmp/uv-cache-part-selection.txt\"\n if ! RELEASE_JSON=\"${release_json}\" PART_PREFIX=\"${part_prefix}\" python3 -c \"import json, os, re, sys; payload=json.loads(os.environ['RELEASE_JSON']); part_prefix=os.environ['PART_PREFIX']; pattern=re.compile(r'^' + re.escape(part_prefix) + r'(\\\\d{3})$'); parts=[]; [parts.append((int(m.group(1)), int(a.get('id')), a.get('name'))) for a in payload.get('assets', []) for m in [pattern.match(a.get('name', ''))] if m and a.get('id') is not None]; parts.sort(key=lambda x: x[0]); indices=[p[0] for p in parts]; expected=list(range(len(parts))); print('\\\\n'.join(f'{asset_id} {name}' for _, asset_id, name in parts)) if parts and indices == expected else (_ for _ in ()).throw(SystemExit(2 if not parts else 3))\" > \"${part_selection_file}\"; then\n echo \"::error::No complete uv cache part set found for prefix '${part_prefix}'.\"\n exit 1\n fi\n\n part_count=\"$(wc -l < \"${part_selection_file}\" | tr -d ' ')\"\n echo \"Using uv cache part set '${part_prefix}*' (${part_count} parts).\"\n\n parts_dir=\"/tmp/uv-cache-parts\"\n part_paths_file=\"/tmp/uv-cache-part-paths.txt\"\n rm -rf \"${parts_dir}\"\n mkdir -p \"${parts_dir}\"\n awk -v d=\"${parts_dir}\" '{print d \"/\" $2}' \"${part_selection_file}\" > \"${part_paths_file}\"\n\n PARTS_DIR=\"${parts_dir}\" GITHUB_TOKEN=\"${GITHUB_TOKEN}\" GITHUB_REPOSITORY=\"${GITHUB_REPOSITORY}\" \\\n xargs -n 2 -P 8 sh -c '\n asset_id=\"$1\"\n asset_name=\"$2\"\n part_path=\"${PARTS_DIR}/${asset_name}\"\n curl -fsSL -L \\\n -H \"Authorization: Bearer ${GITHUB_TOKEN}\" \\\n -H \"Accept: application/octet-stream\" \\\n \"https://api.github.com/repos/${GITHUB_REPOSITORY}/releases/assets/${asset_id}\" \\\n -o \"${part_path}\"\n ' sh < \"${part_selection_file}\"\n\n while IFS= read -r part_path; do\n [ -s \"${part_path}\" ] || {\n echo \"::error::Missing or empty cache part: ${part_path}\"\n exit 1\n }\n done < \"${part_paths_file}\"\n\n rm -rf \"${UV_CACHE_DIR}\"\n mkdir -p \"${UV_CACHE_DIR}\"\n while IFS= read -r part_path; do\n cat \"${part_path}\"\n done < \"${part_paths_file}\" | zstd -d -c | tar -xf - -C \"${UV_CACHE_DIR}\"\n du -sh \"${UV_CACHE_DIR}\"\n\n - name: Install dependencies (with all optional extras for complete type checking)\n run: |\n py_mm=\"$(python -c 'import sys; print(f\"{sys.version_info.major}.{sys.version_info.minor}\")')\"\n cudnn_path=\"${GITHUB_WORKSPACE}/.venv/lib/python${py_mm}/site-packages/nvidia/cudnn\"\n export CUDNN_PATH=\"${cudnn_path}\"\n export CUDNN_HOME=\"${cudnn_path}\"\n export CUDNN_INCLUDE_PATH=\"${cudnn_path}/include\"\n export CUDNN_LIBRARY_PATH=\"${cudnn_path}/lib\"\n export CPLUS_INCLUDE_PATH=\"${CUDNN_INCLUDE_PATH}${CPLUS_INCLUDE_PATH:+:${CPLUS_INCLUDE_PATH}}\"\n export LIBRARY_PATH=\"${CUDNN_LIBRARY_PATH}${LIBRARY_PATH:+:${LIBRARY_PATH}}\"\n export LD_LIBRARY_PATH=\"${CUDNN_LIBRARY_PATH}${LD_LIBRARY_PATH:+:${LD_LIBRARY_PATH}}\"\n uv --version\n uv sync --all-extras --group dev --frozen\n\n - name: Run prek hooks (lint, format, typecheck, uv.lock, tests)\n run: |\n uv run prek run --all-files\n\n - name: Run unit tests (via prek)\n run: |\n uv run prek run pytest\n" }, { "path": ".github/workflows/release.yml", "content": "name: Release\n\non:\n pull_request:\n types: [closed]\n branches: [main]\n\npermissions:\n contents: write\n id-token: write\n\njobs:\n release:\n runs-on: ubuntu-latest\n if: github.event.pull_request.merged == true && startsWith(github.event.pull_request.head.ref, 'release/')\n steps:\n - uses: actions/checkout@v4\n with:\n fetch-depth: 0\n\n - name: Set up Python\n uses: actions/setup-python@v5\n with:\n python-version: '3.11'\n\n - name: Install uv\n run: |\n curl -LsSf https://astral.sh/uv/install.sh | sh\n echo \"$HOME/.cargo/bin\" >> $GITHUB_PATH\n\n - name: Install dependencies\n run: |\n uv venv\n uv pip install -e .\n uv pip install hatch\n\n - name: Build package\n run: uv run hatch build\n\n - name: Get version from pyproject.toml\n id: get_version\n run: |\n VERSION=$(python -c \"import tomllib; print(tomllib.load(open('pyproject.toml', 'rb'))['project']['version'])\")\n echo \"VERSION=$VERSION\" >> $GITHUB_OUTPUT\n\n - name: Create git tag\n run: |\n git config --local user.email \"action@github.com\"\n git config --local user.name \"GitHub Action\"\n git tag v${{ steps.get_version.outputs.VERSION }}\n git push origin v${{ steps.get_version.outputs.VERSION }}\n\n - name: Publish draft release\n env:\n GH_TOKEN: ${{ github.token }}\n run: |\n # Check if draft release exists and publish it\n if gh release view v${{ steps.get_version.outputs.VERSION }} --json isDraft | jq -r '.isDraft' | grep -q true; then\n gh release edit v${{ steps.get_version.outputs.VERSION }} --draft=false\n else\n echo \"::error::No draft release found for v${{ steps.get_version.outputs.VERSION }}\"\n exit 1\n fi\n\n - name: Upload assets to release\n env:\n GH_TOKEN: ${{ github.token }}\n run: |\n gh release upload v${{ steps.get_version.outputs.VERSION }} dist/*\n\n - name: Publish to PyPI\n uses: pypa/gh-action-pypi-publish@release/v1\n with:\n password: ${{ secrets.PYPI_ART_TOKEN }}\n" }, { "path": ".gitignore", "content": "__pycache__/\n.art/\n.env\n.venv/\ngrpo_trainer_lora_model/\nlogs/\nshared_cache.db\ndata/cache.db\nstreaming-chat-completions/\nunsloth_compiled_cache/\nwandb/\ndocs/node_modules/\ndist/\nreplays/\ntrajectories/\n.DS_Store\n.local/\n.claude/settings.local.json\n.vscode/\n.ruff_cache/\n!/src/art/wandb/\n!/src/art/wandb/**\n/src/art/wandb/__pycache__/" }, { "path": ".pre-commit-config.yaml", "content": "repos:\n - repo: https://github.com/astral-sh/ruff-pre-commit\n rev: v0.12.1\n hooks:\n - id: ruff\n - id: ruff-format\n\n - repo: local\n hooks:\n - id: ty\n name: ty type checking\n entry: uv run ty check src tests\n language: system\n pass_filenames: false\n\n - id: uv-lock-check\n name: uv.lock sync check\n entry: uv lock --check\n language: system\n pass_filenames: false\n\n - id: pytest\n name: Unit tests (manual)\n entry: uv run pytest --nbval --current-env --tb=short tests/unit\n language: system\n pass_filenames: false\n stages: [manual]\n" }, { "path": ".python-version", "content": "3.11\n" }, { "path": ".skyignore", "content": "__pycache__/\n.art/\n# .env\n.venv/\ngrpo_trainer_lora_model/\nlogs/\nshared_cache.db\nstreaming-chat-completions/\nunsloth_compiled_cache/\nwandb/\ndocs/node_modules/\ndist/\ndev/art-e/data/\nreplays/\ntrajectories/\n.DS_Store\n# .local/" }, { "path": "AGENT.md", "content": "## uv package manager by default\n\nThis project uses the `uv` package manager.\n\n- To add a dependency, run `uv add `.\n- To run a script, run `uv run