# Training Data

Place rendered SFT JSONL at `coding_sft.jsonl`.

Expected minimal format for `scripts/train_sft_lora.py`:

```jsonl
{"text": "<full rendered chat transcript ending with the assistant answer>"}
```

For tool-use or multi-turn training, keep a raw copy with structured messages
and generate this rendered file as a build artifact so template changes are
auditable.

Raw message format:

```jsonl
{"id":"example-1","license":"mit","messages":[{"role":"user","content":"Write add(a, b)."},{"role":"assistant","content":"def add(a, b):\n    return a + b"}]}
```

Render and validate:

```bash
.venv-train/bin/python scripts/render_sft_jsonl.py \
  --input data/train/coding_sft_messages.jsonl \
  --output data/train/coding_sft.jsonl

.venv-train/bin/python scripts/validate_sft_jsonl.py \
  --train-jsonl data/train/coding_sft.jsonl \
  --eval-jsonl data/eval/coding_sft_eval.jsonl
```

Stopped-run repair artifacts:

- `coding_sft_repair_messages.jsonl`: synthetic strict-output repair messages.
- `coding_sft_code_only_messages.jsonl`: code-only rows extracted from fenced
  production answers.
- `coding_sft_prod_repair_messages.jsonl`: 40,000 production message rows plus
  4,096 strict repair rows.
- `coding_sft_prod_repair.jsonl`: rendered training split for a future resumed
  run.
- `coding_sft_prod_repair_strong_messages.jsonl`: recommended local repair
  blend with 24,000 production rows, 16,000 extracted code-only rows, and 4,096
  strict repair rows.
- `coding_sft_prod_repair_strong.jsonl`: rendered recommended repair split for
  a future resumed run.

Do not launch SFT from these files until the user explicitly resumes training.