CooperBench/qwen35-9b-async-coop

What this is

Cooperative two-agent coding dataset: 50 task pairs across 15 repos (random-50 subset), generated with mini_swe_agent on Qwen/Qwen3.5-9B in coop setting using an async prompt variant — agents work fully asynchronously without active coordination (no messaging, no synchronisation points). Patches are auto-merged after both submit. All 50 pairs were successfully evaluated.

At a glance

Field	Value
Model	Qwen/Qwen3.5-9B
Agent	mini_swe_agent (async prompt — no coordination)
Setting	coop
Subset	random-50 (50 completed)
Repos	15
Pairs evaluated	50
Both-pass	2.0% (1/50)
Per-feature pass	15.0% (15/100)
Merge clean rate	64.0% (32/50)
Merge conflict rate	36.0% (18/50)
Error exits	0
Total tokens	~15.7M (in+out, from traj files)
Owner	Arya Prabhudesai
Date	2026-06-03

How it was generated

cooperbench run --setting coop -a mini_swe_agent -c 30 qwen35-9b-async-coop

Model served via vLLM OpenAI-compatible endpoint (openai/Qwen/Qwen3.5-9B). Async variant: agents are not prompted to coordinate at all — they work in parallel without messaging or check-ins. This serves as a useful baseline showing what happens with zero coordination overhead.

File layout

• index.csv — one row per task pair; HF Dataset Viewer entry point
• qwen35-9b-async-coop/coop/<repo>/<task_id>/<features>/ — raw per-pair artifacts: result.json, eval.json, agent1_traj.json, agent2_traj.json, agent{1,2}.patch, conversation.json

log_dir column in index.csv points to the per-pair subdirectory.

Schema highlights for mid-training

Filter on: both_passed=true, model, agent_framework.

metadata JSON carries per-agent statuses, steps, merge outcome, per-feature pass — use json.loads(row["metadata"]) without following the pointer.

Note: total_tokens is 0 for this run — token counts are in agent_full_traj.json under messages[*].extra.response.usage (~15.7M total in+out, lowest of random-50 variants due to no messaging overhead).

Caveats

• random-50 subset — all 50 tasks completed
• 9.0% LimitsExceeded exits (9/100 agent slots); 0 Error exits
• 36.0% merge conflict rate (18/50)
• Token fields in result.json are 0; aggregate from agent_full_traj.json if needed

Citation

@dataset{qwen35_9b_async_coop,
  title  = {qwen35-9b-async-coop},
  author = {Arya Prabhudesai},
  year   = {2026},
  url    = {https://huggingface.co/datasets/CooperBench/qwen35-9b-async-coop}
}