feat: traces -> hf dataset converter

[mikasenghaas]

Summary

• Replace scripts/export_sft.py with tools/convert_traces_to_hf_dataset.py.
• Convert each branch of each agent trace into one dataset row.
• Export trainable and non-trainable agents without selection or filtering.
• Add top-level outcome fields for downstream filters, including reward, stop_condition, has_error, and is_truncated.
• Preserve task, run, agent, error, timing, usage, call, reward, metric, and trace metadata.
• Push to a private Hugging Face Hub dataset by default, with --public as an explicit opt-out.
• Use --local to write Parquet files with Hugging Face dataset-card metadata.
• Preserve existing dataset-card content when local metadata is updated.

Verification

Eval → dataset → SFT

1. Generated eight GSM8K episodes without uploading the eval:

uv run --project /home/ubuntu/prime-rl eval gsm8k -n 8 -r 1 -c 8 \
    --no-push -o /tmp/trace-converter-e2e-20260825 \
    --run.name trace-converter-gsm8k --run.dir eval

All eight episodes succeeded. Mean reward was 1.0 and the error rate was 0.0.

2. Converted the trace file into a local Hugging Face dataset:

uv run --project /home/ubuntu/prime-rl python \
    tools/convert_traces_to_hf_dataset.py \
    /tmp/trace-converter-e2e-20260825/eval/traces.jsonl \
    --name /tmp/trace-converter-e2e-20260825/dataset --local

The converter wrote eight branch rows. The dataset loaded through load_dataset.

3. Trained Qwen3-0.6B on the converted dataset:

uv run --project /home/ubuntu/prime-rl sft \
    --output-dir /tmp/trace-converter-e2e-20260825 --run.name sft \
    --max-steps 5 --model.name Qwen/Qwen3-0.6B \
    --data.type sft --data.name /tmp/trace-converter-e2e-20260825/dataset \
    --data.subsets '["default"]' --data.splits '["train"]' \
    --data.seq-len 2048 --data.batch-size 4 --data.micro-batch-size 1 \
    --deployment.gpus-per-node 2 --deployment.num-train-gpus 1

The run completed all five steps. Final loss was 0.7764, the NaN count was zero, and peak GPU memory was 10.7 GiB. No Hub or Prime platform uploads occurred.

Additional checks

• Ran the Ruff check and format pre-commit hooks.
• Converted a separate saved file with 132 episodes and loaded all 132 branch rows with load_dataset.
• Verified that a non-trainable errored trace still emits its branch row and top-level error fields.
• Verified that local updates preserve plain, valid-frontmatter, and unclosed-frontmatter README content.
• Verified that default uploads pass private=True, while --public passes private=False.

Breaking

• scripts/export_sft.py is removed. Use uv run python tools/convert_traces_to_hf_dataset.py <traces.jsonl> --name <dir-or-repo-id>.
• The converter no longer accepts --min-reward or --drop-truncated. Apply filters in a separate script.
• The converter pushes to the Hub by default. Use --local to write a local dataset.
• New Hub datasets are private by default. Use --public to create a public dataset.
• The --push flag is removed because Hub upload is now the default.

🤖 Generated with Claude Code

---

Note

Low Risk
Changes are limited to offline export tooling and CLI workflows; no in-repo training or runtime code is modified in this diff.

Overview
Removes scripts/export_sft.py and adds tools/convert_traces_to_hf_dataset.py as the path from eval/train traces.jsonl to Hugging Face datasets.

The new converter reads one episode per line (WireEpisode) instead of per-trace wire rows, emits one row per branch for every agent trace (no --min-reward, --drop-truncated, or empty-branch skipping), and keeps messages plus JSON tools for SFT while adding top-level filter fields (reward, stop_condition, has_error, is_truncated, etc.) and JSON metadata for task, run, agent, errors, timing, and usage.

Output behavior changes: Hub upload is the default via required --name; use --local to write <name>/<subset>/<split>.parquet and update README.md dataset-card configs for load_dataset. --public controls visibility on new Hub repos; --push is gone.

^{Reviewed by Cursor Bugbot for commit 8834b7a. Bugbot is set up for automated code reviews on this repo. Configure here.}

[cursor]

Cursor Bugbot has reviewed your changes and found 2 potential issues.

^{❌ Bugbot Autofix is OFF. To automatically fix reported issues with cloud agents, enable autofix in the Cursor dashboard.}

^{Reviewed by Cursor Bugbot for commit 08c1be0. Configure here.}