[akanyaani]

Fixes from a first-time-user review: install from pip in an empty folder, then follow the README.

Blockers fixed

• Recipes now ship in the package. decisionformer train deciformer15k_qwen3 works after a plain pip install; before, configs/recipes/... only existed in a clone. Related commands:
  • decisionformer recipes lists the recipes.
  • decisionformer init NAME copies one to edit.
  • A config can inherit a recipe with base: NAME.
• The starting recipe learns. The new deciformer15k_qwen3 (Qwen3-0.6B + LoRA) replaces the Qwen3.5 example, which has not learned reliably. The Qwen3.5 recipe is removed until it does; details are in docs/benchmarks.md.
• predict reads every accepted record shape: requests, single questions and flat rows such as deciformer-15k. It batches requests, and a record it can't answer becomes an error line instead of crashing the run.
• README install section: it says to install from GitHub until the first PyPI release, since PyPI currently only holds the placeholder.

Friction fixed

• Safe defaults in code. configs/base.yaml is gone and its values are now the code defaults:
  • Resumable state and evaluation every 500 steps, with resume: auto and max_length 1024.
  • The learning rate follows the fine-tuning mode: 2e-5 full, 2e-4 LoRA/QLoRA, 1e-3 head-only.
  • Resumable state is deleted when a run finishes; it was 1.6 GB for the laptop recipe.
  • Rerunning a finished run warns that it will overwrite it.
• Python training API. decisionformer.train(...) takes a recipe name, a YAML file or a dict, plus dotted overrides. load_config accepts the same.
• Logs.
  • The Hub's httpx2 request lines are silenced.
  • The precision is named ("float32", "fp16 autocast with loss scaling").
  • Each step shows step N/total and the time left.
  • Training warns when the training data shares ids, families or identical states with the dev or calibration data.
• Server.
  • Requests that arrive while the model is busy are answered together, up to --max-batch (default 16); a single request still runs alone.
  • A failing request in a batch fails alone.
  • /docs now shows the request and response schemas.
• Model names. A local runs/<name>/best answers as <name>, not best.
• CLI help shows examples and points to the recipes.

Testing

All on the GPU box; nothing was run on a laptop.

• Full suite with the GPU hidden: 265 passed, 12 skipped. New tests cover recipe names, dict configs, the learning rate by mode, recipes/init, predict with mixed shapes and a bad row, server batching with error isolation, the /docs schema, the leakage warning, train() from Python, and the state cleanup after pause/resume.
• Pip-user simulation: a fresh venv with CPU torch, the built wheel and an empty folder. recipes, init, data split, train local_smollm2 (3 steps), predict (3 rows plus a bad one), serve, /docs, and 32 requests from 16 concurrent clients all work. The training log has no httpx lines.
• deciformer15k_qwen3, one epoch on the RTX 2060 SUPER (fp16 fallback): dev accuracy 0.792 → 0.805 → 0.824 at steps 125/250/375, calibrated ECE 0.023, per task 0.75–0.90.
• Qwen3.5 (bidirectional, QLoRA, last-token readout), 40-step check: the export is the correct Decision 3.0 layout, but the loss spiked and dev accuracy was 0.41 on 128 rows, so no Qwen3.5 recipe ships.

Not in this PR

• A Colab notebook.
• Typed Python response objects.
• A real run on Apple-silicon (MPS) hardware.