asaverren/openhands-train-ready

openhands-train-ready

SFT-ready normalization of nebius/SWE-rebench-openhands-trajectories — 5,000 resolved OpenHands (v0.54.0) agent trajectories on real GitHub issues, bootstrapped by Qwen/Qwen3-Coder-480B-A35B-Instruct, with tool-call arguments deserialized to objects so chat templates just work.

What was fixed

The upstream parquet stores tool_call.function.arguments as a serialized JSON string "for storage efficiency" (per the Nebius README). Chat templates that iterate arguments|items — including the Qwen3-Coder family — emit empty/broken tool calls if you train on the raw rows. Every record here has:

• —arguments deserialized str → dict (double-encoded JSON and Python-repr dicts handled),
• —canonical per-role fields (system/user/assistant/tool),
• —the per-trajectory tools list kept on the record — pass it straight to apply_chat_template.

Usage

from datasets import load_dataset
from transformers import AutoTokenizer

ds = load_dataset("asaverren/openhands-train-ready", split="train")
tok = AutoTokenizer.from_pretrained("Qwen/Qwen3-Coder-30B-A3B-Instruct")

ex = ds[0]
text = tok.apply_chat_template(ex["messages"], tools=ex["tools"], tokenize=False)

TRL/axolotl: point your chat-template mapping at the messages column; tools is a per-row column for templates that accept it.

Qwen3-Coder tool-call format (verified)

• —Assistant calls render as <tool_call><function=NAME><parameter=KEY>VALUE </parameter></function></tool_call> inside the <|im_start|>assistant turn.
• —arguments must be a mapping — the template iterates arguments|items. (This is the upstream storage bug this dataset fixes.)
• —Tool results render as <tool_response>...</tool_response> inside a user turn; name/tool_call_id are kept for provenance but unused by the template.

Counts (see meta.json)

Tool calls validated against upstream tools.json (soft; all calls match the 5-tool OpenHands v0.54.0 set). All 5,000 records render through the real Qwen3-Coder chat template with zero errors.

Rebuild / extend

run_normalize.py (+ the traj_normalize package, both in this repo) reproduces this shard and can emit the full 32k resolved set (or failure trajectories via --include-failures):

pip install pyarrow huggingface_hub
python run_normalize.py --input nebius/SWE-rebench-openhands-trajectories \
  --out-dir out --max-success 5000 --seed 0 --shuffle

sft_success.jsonl.gz is the raw tool output; data/ is the same content as parquet for load_dataset + viewer support.

Honest limitations

• —resolved is upstream's SWE-bench-style judgment — no re-execution verification was performed downstream.
• —Trajectories are long (~49k tokens avg); filter/truncate to your context budget.
• —Single bootstrap model (Qwen3-Coder-480B) and single scaffold (OpenHands v0.54.0) — behavior/diversity is bounded accordingly.
• —Same-instance duplicates are possible (upstream samples multiple attempts per issue); dedupe on instance_id if that matters to you.

Attribution & license

Upstream data: Nebius, CC-BY-4.0. This derivative shard redistributes normalized data under the same license — please attribute Nebius.

@article{trofimova2025openhandstrajs,
  title={OpenHands Trajectories with Qwen3-Coder-480B-A35B-Instruct},
  author={Trofimova, Maria and Shevtsov, Anton and Ibragim, Badertdinov and Pyaev, Konstantin and Karasik, Simon and Golubev, Alexander},
  year={2025},
  journal={Nebius blog},
}