cwz19/cooper-solo2coop_fail

Cooper solo→coop FAILURES (Qwen3.5-9B, synthesized)

Two-agent coop coding trajectories synthesized from FAILED single-agent solo CooperBench runs via coopertrain.solo_to_coop. Each solo run (one model implementing two features, which did not pass both feature tests) is split into two cooperating agents that own one feature each and coordinate via send_message. Every row has both_passed=False — this is negative / attempt data (useful for RL, contrastive training, or studying failure modes), not gold demonstrations. Coordination is failure-aware (no false claims of success).

At a glance

Model	Qwen3.5-9B
Agent framework	mini_swe_agent (coop)
Source	CooperBench 99-task manifest (FAILED solo runs; one per pair)
Trajectories (pairs)	832
Repos covered	38
Total tokens	28,012,922 (~0.028B, tiktoken cl100k approx)
Source both-passed	0/832
Own-patch decomposed	190/832
LLM-rewritten	832/832
Deferral exchanges	188

How it was generated

# 1. Solo rollout (Qwen3.5-9B on Modal vLLM) over all feature pairs
python scripts/batch_collect_9b.py --name qwen35-9b-batch-v1 --concurrency 8

# 2. Solo -> coop synthesis (per-feature patch decompose + thinking rewrite, Flash)
uv run python -m coopertrain.solo_to_coop \
    --input logs/qwen35-9b-batch-v1/ --output logs/qwen35-9b-batch-v1/coop/ \
    --decompose-patches --llm-rewrite --model gemini/gemini-2.5-flash --temperature 0.3

# 3. Export canonical coop run-dir format (agent{fid}_traj.json + patches + conversation + result)
uv run python scripts/export_coop_canonical.py \
    --coop-root logs/qwen35-9b-batch-v1/coop --dataset-root dataset --model openai/Qwen/Qwen3.5-9B

File layout

• index.csv — one row per coop pair (the Dataset Viewer entry point): repo, task_id, features, setting, agent_framework, model, both_passed, tokens, n_messages, log_dir, metadata. log_dir points into the raw tree; metadata is a JSON blob (per-agent steps/patch_lines, breakpoints, labels, decompose/rewrite flags).
• Raw tree — <repo>/task<id>/<features>/ per pair, in canonical coop format (matches a live cooperbench run --setting coop):
  • agent{fid}_traj.json — that agent's litellm chat trajectory (system/user/assistant+tool_calls/tool); the trainable artifact.
  • agent{fid}.patch — that agent's per-feature unified diff.
  • conversation.json — inter-agent send_message exchanges.
  • result.json — per-agent status/steps wrapper.
  • readable_merged.json — merged viewer timeline (companion, not for training).
  • solo_to_coop_metadata.json — synthesis provenance.

Schema highlights for mid-training

• both_passed is False for every row — failed solo attempts.
• Stratify by repo / model. tokens is the per-pair size signal.
• Train per agent: each pair yields two SFT examples (one per agent{fid}_traj.json); see scripts/prepare_sft_data.py.

Caveats

• Synthesized, not real coop rollouts. Coordination (planning/kickoff/ midedit/farewell send_message) is LLM-synthesized (Gemini Flash) from a solo trajectory; per-feature patches are split from the agent's own solo patch (not dataset gold patches). Token counts are tiktoken cl100k approximations.
• A small number of source pairs where the classifier found no feature-specific moves and patch decomposition failed were dropped (no coordination signal).

Citation

@misc{cooper_solo2coop_succ,
  title  = {Cooper solo->coop (Qwen3.5-9B, synthesized)},
  author = {Chen, Wentse},
  year   = {2026},
  howpublished = {\url{https://huggingface.co/datasets/cwz19/cooper-solo2coop_fail}}
}