datasets
Training and evaluation data, with the modality, task and licence stated up front. Listed live from the Hugging Face Hub.
cooperbench-dataset
CooperBench Dataset
This dataset contains the benchmark tasks for evaluating multi-agent coordination in code collaboration.
Run it with the official harness: github.com/cooperbench/CooperBench (pip install cooperbench, then cooperbench prepare downloads this dataset). The harness, the Docker images, and this dataset are versioned together — use the GitHub repo's dataset/ tree or this mirror at the matching tag.
Paper: CooperBench: Why Coding Agents Cannot be Your Teammates Yet… See the full description on the dataset page: https://huggingface.co/datasets/CooperBench/cooperbench-dataset.trajectoriescooperagents-programbench-tracesteam-coop
CooperBench Team → Coop (codex / gpt-5.5-hao, full dataset)
Two complete cooperbench team-harness runs (lead + member sharing a
Redis-backed task list, scratchpad volume, and MCP) reshaped into the
2-agent coop layout defined in cooperbench/CooperData PR
#98 and produced by
cooperbench/CooperData PR #101
(cooperdata.cooperation.ingest.convert_team).
run
pairs
pass-rate
setting
source tarball
cmp-full-team-noproto
652
61.8% (403/652)
task_list + scratchpad + mcp +… See the full description on the dataset page: https://huggingface.co/datasets/CooperBench/team-coop.team-trajectories
CooperBench Coordination Study — Agent Trajectories
Raw agent trajectories from the CooperBench coordination comparison: how
multi-agent coordination affects task success when two agents split a
two-feature task.
Agents: codex (OpenAI Codex CLI) and mini_swe_agent_v2
Model: gpt-5.5-hao (Azure OpenAI deployment)
Backend: docker · seed: 1
Files
Each tarball expands to <run>/<setting>/<repo>/<task>/<f_a>_<f_b>/ task
directories. Per task: agentN_stream.log /… See the full description on the dataset page: https://huggingface.co/datasets/CooperBench/team-trajectories.cooperator-sft-dataqwen9b-coop-claude-code
qwen9b-coop-claude-code
Two-agent cooperative coding trajectories generated by running
CooperBench in coop mode on
the CooperData task set, using
Qwen/Qwen3.5-9B as the model and Claude Code (claude_code) as the
agent framework. Each pair runs two agents in parallel — one per feature —
coordinating via Redis messaging and a shared git remote.
The matched solo (single-agent) baseline is at
CooperBench/qwen9b-solo-claude-code.
Same task corpus, same model, same agent — only the… See the full description on the dataset page: https://huggingface.co/datasets/CooperBench/qwen9b-coop-claude-code.qwen9b-solo-claude-code
qwen9b-solo-claude-code
Single-agent coding trajectories generated by running
CooperBench in solo mode on
the CooperData task set, using
Qwen/Qwen3.5-9B as the model and Claude Code (claude_code) as the
agent framework. One agent implements both features in each task.
The matched coop (two-agent) version is at
CooperBench/qwen9b-coop-claude-code.
Same task corpus, same model, same agent — only the coordination differs, so
together they isolate the cooperation deficit.
At a… See the full description on the dataset page: https://huggingface.co/datasets/CooperBench/qwen9b-solo-claude-code.cooperdata-sft-midtrain
CooperData — bucketed (SFT / mid-training / discarded)
Unified, coordination-quality-bucketed view of the CooperBench cooperative coding-agent datasets. Each row is one coop pair (two agents each implementing a feature in a shared repo), normalized to a common schema with a source column.
Splits
split
bucket
meaning
rows
sft
A
exemplary coordination workflow worth imitating
2455
midtraining
B
coordination present but thin / one-sided / synthetic
3849… See the full description on the dataset page: https://huggingface.co/datasets/CooperBench/cooperdata-sft-midtrain.qwen35-9b-cooper-midtrain-trajectories
Qwen3.5-9B cooper-midtrain lineage — CooperBench coop trajectories
Agent trajectories from the July 2026 evaluation of the cooper midtrain
lineage (Qwen3.5-9B-cooper-{instruct-midtrain,midtrain-sft,midtrain}-merged)
on CooperBench coop tasks. Stock coop prompt, mini_swe_agent_v2 harness,
auto-eval on. Produced on the feature/opsd-training evidence trail
(see coopertrain/opsd/evidence/EXPERIMENT_LOG.md).
folder
model
scan
pairs
step limit
date
cooper-imt-v1-stock… See the full description on the dataset page: https://huggingface.co/datasets/CooperBench/qwen35-9b-cooper-midtrain-trajectories.cooperdata-v3-midtrain-blend
CooperData v3 — Midtraining Blend (Qwen3.5-9B cooperative SWE agents)
All-token midtraining mixture that bridges Qwen/Qwen3.5-9B (instruct) toward the cooperative
multi-agent SWE-coding SFT distribution. One document per row (text, tagged by source) — NOT
packed — so trl.SFTTrainer(packing=False) tokenizes per-doc and the Gated-DeltaNet recurrence
stays per-document. ~390M tokens.
Composition
source
tokens
share
role
web
210.0M
54%
general
math
55.0M… See the full description on the dataset page: https://huggingface.co/datasets/CooperBench/cooperdata-v3-midtrain-blend.qwen9b-coop-mini-swe-agent
qwen9b-coop-mini-swe-agent
Two-agent cooperative coding trajectories generated by running
CooperBench in coop mode on
the CooperData task set, using
Qwen/Qwen3.5-9B as the model and mini_swe_agent_v2 as the agent framework.
Each pair runs two agents in parallel — one per feature — coordinating via Redis messaging and a shared git remote.
The matched solo version is at
CooperBench/qwen9b-solo-mini-swe-agent.
Same task corpus, same model, same agent — only the coordination differs… See the full description on the dataset page: https://huggingface.co/datasets/CooperBench/qwen9b-coop-mini-swe-agent.qwen9b-solo-mini-swe-agent
qwen9b-solo-mini-swe-agent
Single-agent coding trajectories generated by running
CooperBench in solo mode on
the CooperData task set, using
Qwen/Qwen3.5-9B as the model and mini_swe_agent_v2 as the agent framework.
One agent implements both features in each task.
The matched coop version is at
CooperBench/qwen9b-coop-mini-swe-agent.
Same task corpus, same model, same agent — only the coordination differs, so
together they isolate the cooperation deficit.
At a glance… See the full description on the dataset page: https://huggingface.co/datasets/CooperBench/qwen9b-solo-mini-swe-agent.qwen35-9b-plan-first-coop
What this is
Cooperative two-agent coding dataset: 211 task pairs across 18 repos, generated with
mini_swe_agent on Qwen/Qwen3.5-9B in coop setting using a plan-first prompt variant —
agents are prompted to produce an explicit implementation plan before writing code, then coordinate
to reconcile plans before proceeding. Patches are auto-merged after both submit.
At a glance
Field
Value
Model
Qwen/Qwen3.5-9B
Agent
mini_swe_agent (plan-first prompt)… See the full description on the dataset page: https://huggingface.co/datasets/CooperBench/qwen35-9b-plan-first-coop.qwen35-9b-git-coop
What this is
Cooperative two-agent coding dataset: 211 task pairs across 18 repos, generated with
mini_swe_agent on Qwen/Qwen3.5-9B in coop setting with a shared read-only git remote (--git).
Agents coordinate via messaging and git fetch team; patches are auto-merged after both submit.
At a glance
Field
Value
Model
Qwen/Qwen3.5-9B
Agent
mini_swe_agent (step_limit=300)
Setting
coop + git remote
Repos
18
Pairs
211
Both-pass
5.7% (12/210… See the full description on the dataset page: https://huggingface.co/datasets/CooperBench/qwen35-9b-git-coop.cooperdata-sft-midtrain-v2
CooperData v2 — contamination-free, coordination-quality-bucketed
Unified view of the CooperBench cooperative coding-agent datasets (+ cooperative-game logs), one row per coop pair, for training a 9B model to be better at CooperBench.
Train/test safety: every pair whose (repo, task_id) is one of the 30 held-out CooperBench benchmark tasks is hard-excluded (X) before bucketing — zero benchmark leakage. team-trajectories and the codex team-coop/cmp-full-team* arms were 100% on… See the full description on the dataset page: https://huggingface.co/datasets/CooperBench/cooperdata-sft-midtrain-v2.cooperdata-bridge2x-midtrain-blend
CooperData bridge2x — Midtraining Blend (Qwen3.5-9B cooperative SWE agents)
All-token midtraining mixture (recipe bridge2x) that bridges Qwen/Qwen3.5-9B (instruct)
toward the cooperative multi-agent SWE-coding SFT distribution. One document per row (text,
tagged by source) — NOT packed — so trl.SFTTrainer(packing=False) tokenizes per-doc and the
Gated-DeltaNet recurrence stays per-document. ~200M tokens.
Composition
source
tokens
share
role
coop
120.1M… See the full description on the dataset page: https://huggingface.co/datasets/CooperBench/cooperdata-bridge2x-midtrain-blend.cooperator-dual-policy-sft-data
Cooperator Dual-Policy SFT Dataset
Supervised fine-tuning data for training two specialized cooperative coding agents from split-role trajectory data. Each record is a full agent conversation in Qwen 3.5 inline tool-call format, labelled by the policy role the agent played.
Code: github.com/cooperbench/cooperator
Motivation
Standard cooperative SFT treats both agents symmetrically and trains a single policy. This dataset trains two distinct policies instead:… See the full description on the dataset page: https://huggingface.co/datasets/CooperBench/cooperator-dual-policy-sft-data.swechat-coopqwen35-9b-question-first-coop-random-50
What this is
Cooperative two-agent coding dataset: 49 task pairs across 15 repos (random-50 subset), generated
with mini_swe_agent on Qwen/Qwen3.5-9B in coop setting using a question-first prompt variant —
agents begin by asking each other clarifying questions about their respective features before
starting implementation, aiming to surface integration concerns early. All 49 pairs were
successfully evaluated.
At a glance
Field
Value
Model… See the full description on the dataset page: https://huggingface.co/datasets/CooperBench/qwen35-9b-question-first-coop-random-50.qwen35-9b-explore-plan-coop
What this is
Cooperative two-agent coding dataset: 209 task pairs across 18 repos, generated with
mini_swe_agent on Qwen/Qwen3.5-9B in coop setting using an explore-plan prompt variant —
agents explore the codebase first, then produce an explicit implementation plan, then share and
reconcile plans with their partner before writing code. Patches are auto-merged after both submit.
At a glance
Field
Value
Model
Qwen/Qwen3.5-9B
Agent
mini_swe_agent… See the full description on the dataset page: https://huggingface.co/datasets/CooperBench/qwen35-9b-explore-plan-coop.qwen35-9b-leader-follower-coop
What this is
Cooperative two-agent coding dataset: 39 task pairs across 14 repos (random-50 subset), generated
with mini_swe_agent on Qwen/Qwen3.5-9B in coop setting using a leader-follower prompt variant —
one agent is designated as leader and sets the coordination strategy; the other acts as follower
and adapts its implementation plan accordingly. All 39 pairs were successfully evaluated.
Notable: this variant produced the lowest merge conflict rate (17.9%) of all random-50… See the full description on the dataset page: https://huggingface.co/datasets/CooperBench/qwen35-9b-leader-follower-coop.qwen35-9b-cooper-flash-evals
CooperBench flash_10 evaluation traces
Every run here is the same frozen contract: subset flash_10 (10 pairs / 10 tasks / 9 repos),
agent mini_swe_agent_v2, setting coop + --git, Modal sandboxes, step_limit=200,
max_model_len=49152. Produced by scripts/eval_e2e.py in
CooperTrain. Two runs are comparable only if their
run_meta.json agree — check it rather than assuming.
Each run directory contains, per pair:
file
contents
agentN_traj.json
one agent's trajectory… See the full description on the dataset page: https://huggingface.co/datasets/CooperBench/qwen35-9b-cooper-flash-evals.cooperdata-bridge-midtrain-blend
CooperData bridge — Midtraining Blend (Qwen3.5-9B cooperative SWE agents)
All-token midtraining mixture (recipe bridge) that bridges Qwen/Qwen3.5-9B (instruct)
toward the cooperative multi-agent SWE-coding SFT distribution. One document per row (text,
tagged by source) — NOT packed — so trl.SFTTrainer(packing=False) tokenizes per-doc and the
Gated-DeltaNet recurrence stays per-document. ~100M tokens.
Composition
source
tokens
share
role
coop
60.0M
60%… See the full description on the dataset page: https://huggingface.co/datasets/CooperBench/cooperdata-bridge-midtrain-blend.cooperbench-logsqwen9b-coop-claude-code-compressed
qwen9b-coop-claude-code-compressed-ak
Synthetic compressed cooperative agent trajectories derived from
CooperBench/qwen9b-coop-claude-code.
Each raw pair (two LLM coding agents on overlapping features in the same
repo, communicating via Redis messaging + a team git remote) is condensed
into an idealized version: wasted steps dropped, broken submission rituals
fixed, missing cooperation events (coop-send/coop-broadcast/coop-recv,
git fetch/diff team) inserted where the real pair… See the full description on the dataset page: https://huggingface.co/datasets/CooperBench/qwen9b-coop-claude-code-compressed.qwen35-9b-milestone-checkins-coop
What this is
Cooperative two-agent coding dataset: 211 task pairs across 18 repos, generated with
mini_swe_agent on Qwen/Qwen3.5-9B in coop setting using a milestone-checkins prompt variant —
agents use periodic structured check-ins at predefined milestones to coordinate progress and
surface integration conflicts early. Patches are auto-merged after both submit.
Coverage caveat: Only 146 of 211 pairs were successfully evaluated (65 had eval errors).
The high agent Error rate… See the full description on the dataset page: https://huggingface.co/datasets/CooperBench/qwen35-9b-milestone-checkins-coop.qwen35-9b-contract-first-coop-random-50
What this is
Cooperative two-agent coding dataset: 36 task pairs across 13 repos (random-50 subset), generated
with mini_swe_agent on Qwen/Qwen3.5-9B in coop setting using a contract-first prompt variant —
agents first agree on a shared interface contract (function signatures, data structures, API
boundaries) before independently implementing their respective features. All 36 pairs were
successfully evaluated.
At a glance
Field
Value
Model… See the full description on the dataset page: https://huggingface.co/datasets/CooperBench/qwen35-9b-contract-first-coop-random-50.qwen35-9b-late-sync-coop-random-50
What this is
Cooperative two-agent coding dataset: 48 task pairs across 15 repos (random-50 subset), generated
with mini_swe_agent on Qwen/Qwen3.5-9B in coop setting using a late-sync prompt variant —
agents work independently for most of the task and synchronise only at a late stage before
submission. Patches are auto-merged after both submit. All 48 pairs were successfully evaluated.
At a glance
Field
Value
Model
Qwen/Qwen3.5-9B
Agent
mini_swe_agent… See the full description on the dataset page: https://huggingface.co/datasets/CooperBench/qwen35-9b-late-sync-coop-random-50.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… See the full description on the dataset page: https://huggingface.co/datasets/CooperBench/qwen35-9b-async-coop.