whisperle/harnessGRPO-traces

Cross-Harness Trace RL: trace corpus and frozen analysis artifacts

Every trajectory, verdict, and analysis output behind the paper Cross-Harness Trace RL Moves Policies Without Detectable Portable Gains: A Controlled Fixed-Data Study of Coding Agents.

The study asks a narrow question: when reinforcement learning replays trajectories collected by several different agent harnesses (the tools, prompts, retry logic and context management around a model), does the resulting weight update help under a harness the policy never trained on? Four production coding-agent frameworks — Aider, OpenHands, Qwen Code and SWE-agent — collect trajectories on SWE-Gym with Qwen3-8B; the resulting dataset is frozen and replayed once by several RL objectives; every trained checkpoint is then scored against the same base model under the four training harnesses and two held-out ones, using an external hidden-test grader.

The headline result is a bounded null with an unusual amount of machinery behind it, so the corpora are released as traces rather than as summaries: they are what let someone check whether the null is real, or find the structure we missed. The evaluation side is complete; one training-side corpus was reclaimed from cluster storage before release and is documented below rather than quietly omitted.

What is here

Path	Contents	Scale	Download
frozen_results/	Every analysis output the paper's numbers come from	42 files	1.2 MB
task_rosters_and_plan/	Frozen task lists, analysis plan, freeze manifest with hashes	6 files	small
training_logs/	SWE-Gym collection, four harnesses: full per-instance interaction logs	2,401 instances, 128 GB raw	16.3 GB
training_manifests/	The same episodes as RL consumed them: tokens + advantages	74,663 records	0.6 GB
training_inventory/	File inventory and reward checksums for the reclaimed trace corpus	375,577 entries	20 MB
eval_matrix_audit200/	The 8 conditions × 6 harnesses evaluation matrix (audit-200 roster)	~40,000 episodes, 21 GB raw	10.6 GB
probe_e1/	weak-ReAct probe / same-data experiment, k=4	~74k files, 16 GB raw	10.7 GB
base_verified/	Base-model SWE-bench Verified baseline	11,321 files	2.8 GB

Total download: 41 GB across 132 shard and metadata files.

Bulky corpora ship as shard-NNNN.tar.zst archives of roughly 4 GB uncompressed each. The token-level JSON compresses about 16×, so even the largest prefix is a modest download.

# one shard
zstd -d shard-0000.tar.zst -c | tar -xv

# everything under one prefix
huggingface-cli download whisperle/harnessGRPO-traces --repo-type dataset \
  --include 'eval_matrix_audit200/*' --local-dir ./eval_matrix

A note on the training-side traces

The per-episode .captures.json / .traj.json / .patch files from the SWE-Gym collection were reclaimed from cluster storage after the RL manifests were assembled from them, so they are not part of this release. What survives, and is published here, is:

• training_logs/ — the full per-instance interaction logs for all 2,401 SWE-Gym instances (128 GB), which contain the agent-framework side of each episode;
• training_manifests/ — those same episodes in the exact form the RL objectives consumed: turn-level records with task, harness, rollout, turn, input_ids, loss_mask, reward and advantage;
• training_inventory/ — the original file inventory (375,577 entries with sizes) and reward checksums, so the reclaimed corpus's composition and outcomes remain auditable.

The evaluation-side corpora below are complete: nothing was reclaimed from them.

Per-episode file types

Every episode is keyed <instance_id>#<harness>#r<rollout>:

• .traj.json — the trajectory: messages, tool calls, observations. The human-readable artifact, and the one most analyses actually want. Only ~5 % of the corpus by volume.
• .captures.json — token ids and logprobs recorded between the framework and the vLLM engine, so training uses exactly the tokens the engine sampled rather than a re-tokenization. The bulk of the volume.
• .patch — the emitted patch. Includes the pathological tails the paper reports: one SWE-agent episode emitted a 1.74 GB patch after a runaway sed.
• .result.json / .resolved.json — inline and sealed-oracle verdicts. The paper uses the oracle verdicts only; the inline ones are published so the +0.42 pp optimism gap is checkable.

Training manifests are deduplicated

The eight objective streams (within, cross, sft, resid, crossdown, a1, a2, m3) were assembled from one frozen manifest and differ only in the advantage scalar — their input_ids and loss_mask are byte-identical, record for record. Mirroring all eight would ship the same 9.5 GB of tokens eight times, so instead:

• training_manifests/tokens.jsonl.zst — the shared token corpus, once
• training_manifests/advantages.jsonl.zst — one field per objective stream, keyed by task#harness#rollout#turn. Streams are joined by key, not by position: the sft stream covers only the 8,841 successful trajectories it was built from, so that field is present on exactly those records and absent elsewhere — which also makes the field a membership marker for the SFT subset.

Any stream is reconstructed by joining its advantage column onto the token corpus.

Reproducing the paper from this repo

frozen_results/ plus the analysis scripts in the code repository regenerate every table and figure without downloading a single trace:

git clone https://github.com/whisperle/harnessGRPO-iclr2027
huggingface-cli download whisperle/harnessGRPO-traces --repo-type dataset \
  --include 'frozen_results/*' --local-dir .
python scripts/paper_artifacts.py && python scripts/paper_figures_main.py

The traces are needed only to recompute the frozen results themselves, or to ask questions the paper did not.

Provenance and caveats

• Grading. Verdicts come from a sealed per-instance container oracle with four states (resolved / unresolved / invalid_patch / infra_error); infrastructure errors never become zeros and never enter the statistics. Harness-reported verdicts are published but were measured 0.42 pp optimistic and are not used for any reported number.
• Rollout nondeterminism. vLLM continuous batching makes identical reruns flip roughly 6 % of per-task outcomes. Episode seeds are provenance, not reproducibility. No single-run difference should be read as evidence.
• Coverage. 39,811 of 40,000 evaluation-matrix episodes are graded (99.53 %); the ungraded remainder is documented and bounded adversarially in the paper rather than imputed.
• Task split. Training draws on SWE-Gym, evaluation on SWE-bench Verified; the two task sets are disjoint by construction, so held-out gains are end-to-end over new tasks and new harnesses.

Citation

@misc{crossharness2027,
  title  = {Cross-Harness Trace RL Moves Policies Without Detectable Portable Gains:
            A Controlled Fixed-Data Study of Coding Agents},
  year   = {2027},
  note   = {Dataset: https://huggingface.co/datasets/whisperle/harnessGRPO-traces}
}

Released under CC-BY-4.0. Trajectories were generated against public SWE-Gym and SWE-bench Verified instances; the underlying repositories retain their own licenses.