fix(sft): tolerate trace-shape tool calls; opt-in skip for unrenderable samples

[mikasenghaas]

Problem

Plugging a verifiers rollout dataset (e.g. PrimeIntellect/SyntheticData-GLM5.2-scaleswe-reward0, scaleswe_rlm_glm51) into SFT crashes twice over:

1. Immediately, on the first tool-call sample: these datasets store tool_calls entries as flat JSON strings ('{"id", "name", "arguments"}'), and deserialize_tool_calls calls .get on them → AttributeError (src/prime_rl/utils/chat_template.py).
2. On any sample the renderer rejects: SFTDataset.__iter__ calls _process bare, so a renderer ValueError propagates through the dataloader and kills the run. Measured on the dataset above: ~1% of samples are degenerate teacher outputs (garbage tool names like 'ipy'/'ipython<tool_call>ipython…', missing code arg) that a strict renderer correctly refuses.

Change

• deserialize_tool_calls now accepts tool-call entries as JSON strings and normalizes the flat verifiers shape into the OAI {"function": {...}} form (id/type preserved). Dict-shaped OAI entries behave exactly as before.
• New opt-in data.skip_invalid_samples: bool = False: when on, SFTDataset.__iter__ catches ValueError from _process, warns (mirroring the existing zero-trainable-tokens skip log), and continues. Default off — data problems keep failing loudly.

Tests

tests/unit/train/sft/test_sft_dataset.py: trace-shape + OAI-shape deserialization; raising sample crashes with the knob off and is skipped (good samples still yielded) with it on. Full file: 22 passed.

🤖 Generated with Claude Code

---

Note

Low Risk
Localized SFT data preprocessing and an opt-in flag; default behavior stays fail-loud, with no auth or persistence changes.

Overview
Makes SFT usable on verifiers rollout datasets that store tool calls differently and may include a small fraction of unrenderable rows.

deserialize_tool_calls now accepts each tool-call entry as a JSON string or a flat {id, name, arguments} dict, normalizes them to the OAI function shape, and still parses string arguments the same way as before for dict-shaped entries.

data.skip_invalid_samples (default false) wires through to SFTDataset: when enabled, ValueError from rendering/_process is logged and the example is skipped instead of aborting the run; when off, behavior is unchanged (fail fast).

Unit tests cover trace vs OAI deserialization and the skip knob.

^{Reviewed by Cursor Bugbot for commit c9acf6b. Bugbot is set up for automated code reviews on this repo. Configure here.}