These are raw agent trajectories from a benchmark run. They contain verbatim model output, tool calls, and excerpts of source code from the upstream repositories under evaluation. Access is granted for research and evaluation purposes. Please tell us who you are and how you intend to use the data.

SWE-Explore find — dev-100 evidence runs

Five complete 100-trial runs of the find (fault-localization) arm of SWE-Explore, kept because each one is load-bearing evidence for a specific claim about the scoring fixes on branch fix/swe-explore-find-scoring of harness_bench.

Every run is 100 trials of the same dev-100 find subset, run through Harbor with the swe_explore.agents:PiSut agent. index.jsonl has one row per trial (500 rows); the full raw Harbor trial directories are in runs/<run_id>/logs.tar.gz.

The five runs

run_id	arm	code	agent	pi	model	thinking	`reasoning_effort` on the wire	dataset	scored	scorer crashes	zeros	f1 (per scored trial)
`2026-08-28__09-06-36`	qwen before	`242d707`	`PiSut`	0.84.3	`qwen/qwen3.8-27b`	off	`none`	`pool`	98	0	34	0.1424
`2026-08-29__03-24-51`	qwen after	`1466a26`	`PiSut`	0.84.3	`qwen/qwen3.8-27b`	off	`none`	`pool`	98	0	1	0.2086
`2026-08-29__06-18-35`	qwen vs published	`1466a26`	`PiSut`	0.84.4	`qwen/qwen3.8-27b`	off	`none`	published `dev-find`	98	16	18	0.1654
`2026-08-31__13-21-24`	qwen clean	`f07dee7`	`PiSut`	0.84.4	`qwen/qwen3.8-27b`	off	`none`	`pool_clean`	98	0	3	0.2013
`2026-08-31__13-54-05`	muse clean	`f07dee7`	`PiSut`	0.84.4	`meta/muse-glimmer-30b`	minimal	`minimal`	`pool_clean`	98	0	20	0.1488

Common to all five: tools = read,grep,find,ls,write (no bash, no web_search — these are real GitHub issues and egress would let the agent fetch the patch that fixed them), step_limit = 25, environment.type = docker, and OpenRouter as the provider. write is not optional: the answer is a file, not a reply.

Code provenance — read this before comparing numbers

code above is the commit whose tree produced the run. Three of these runs were made from a dirty working tree and the commit landed hours later, so a naive "commit timestamp < run timestamp" check will mislabel them:

run	commit	committed at	run started	state at run time
`2026-08-28__09-06-36`	`242d707`	2026-08-28 07:46	2026-08-28 09:06	committed
`2026-08-29__03-24-51`	`1466a26`	2026-08-29 07:44	2026-08-29 03:24	uncommitted working tree
`2026-08-29__06-18-35`	`1466a26`	2026-08-29 07:44	2026-08-29 06:18	uncommitted working tree
`2026-08-31__13-21-24`	`f07dee7`	2026-08-29 07:45	2026-08-31 13:21	committed
`2026-08-31__13-54-05`	`f07dee7`	2026-08-29 07:45	2026-08-31 13:54	committed

The three commits:

242d707 — fixes three find-mode scoring bugs: the answer channel contradicted itself (instruction said read-only and asked for a file), the vendored scorer could not import under Python < 3.10 (PEP 604 without from __future__ import annotations), and a crashed scorer was recorded as a legitimate reward of 0.
1466a26 — keeps the answer tool alive at the step cap. Past the cap the extension now withdraws the exploration tools only, keeping submit_tools (default ("write",)) for submit_grace_steps turns. Introduced submit_grace_steps, default 3.
f07dee7 — makes thinking_level_map configurable, because meta/muse-glimmer-30b answers reasoning_effort: "none" with 400 Reasoning is mandatory for this endpoint and cannot be disabled.

The code version is recoverable from the logs, not just from the dates. The step cap itself is not logged, but its effect is: reconstruct a capped trial as turn_start events ≥ 26 in agent/pi.txt and check whether the trial left an answer file. That separates the two code versions cleanly:

run	capped	of those, wrote an answer file	mean turns
`2026-08-28__09-06-36`	37/98	4	19.6
`2026-08-29__03-24-51`	37/98	37	18.7
`2026-08-29__06-18-35`	43/98	41	20.4
`2026-08-31__13-21-24`	40/98	40	19.9
`2026-08-31__13-54-05`	48/98	32	22.6

09-06-36 and 03-24-51 have byte-identical agent kwargs — neither passes submit_grace_steps; the after run picks up the new default of 3. So the only difference between them is the code, which is what makes the pair an A/B rather than two runs that happen to differ.

Thinking: off everywhere except the muse run

PiSut defaults --thinking to off and maps levels to reasoning_effort through thinking_level_map, default {"off": "none"}. reasoning_effort: "none" is the protocol this benchmark is measured under, and the four qwen runs all use it.

2026-08-31__13-54-05 cannot. The muse endpoint rejects "none" outright, and its lowest accepted level is minimal, so that run sets thinking: minimal with thinking_level_map: {"minimal": "minimal"}. The two 08-31 runs are therefore not apples-to-apples, and the forced reasoning is not cosmetic: 16 of muse's 48 capped trials spend the grace window emitting a thinking block about needing to write FIND_ANSWER.txt and then end their turn without calling write. That is most of the qwen/muse gap — 19 of muse's 20 zeros wrote no answer file at all, and among the 79 trials that did answer it scores 0.1846 against qwen's 0.2013. Read the headline gap as protocol compliance first, localization skill second.

Dataset differences — this is where the numbers come from

Three distinct task sources appear across the five runs. They are not interchangeable, and the digests recorded in each trial's lock.json prove it:

source	in runs	kind	task digests
`pool`	`09-06-36`, `03-24-51`	local dir (`/tmp/republish/pool`)	identical across both runs, all 100
`pool_clean`	`13-21-24`, `13-54-05`	local dir (`/tmp/republish/pool_clean`)	identical across both runs, all 100; all 100 differ from `pool`
`ai-solution-finetune/swe-explore-dev-find`	`06-18-35`	published package, `sha256:5dfa6483761acec9b1c0aaecc2367512e7182f0dfecda51d17f3a2ac4e337ca0`	published digest

pool and pool_clean are both post-242d707 regenerations, and the agent-visible instruction is byte-identical between them (1932 bytes, verified by diffing the rendered first user message for the same task). All 100 digests differ because pool_clean is a fresh make_harbor_tasks.py run from f07dee7 — a regeneration, not a content change anyone has shown to matter. Treat pool and pool_clean numbers as comparable and pool_clean as the canonical one.
The published swe-explore-dev-find still carries pre-242d707 packages, and that is the whole point of 06-18-35. Its instruction says "Do NOT modify any file, this is a read-only task" and then asks for /testbed/FIND_ANSWER.txt anyway — the contradiction 242d707 removed. Its scorer is also unpatched, which is why 16 of its 18 zeros are the scorer crashing (eval.py:29 TypeError: unsupported operand type(s) for |, PEP 604 under Python 3.9), not the model failing; 15 of those 16 had a valid answer file on disk. Substituting their real scores gives 0.2057, within noise of 03-24-51's 0.2086. Expect 0.1654 from the published dataset until it is republished.
pool_clean was verified byte-identical to what a republish would upload except the name = line in task.toml (local swe-explore/... vs published ai-solution-finetune/...), which no scorer reads.

/tmp/republish/pool and /tmp/republish/pool_clean no longer exist on the run host — /tmp was cleared. The per-trial lock.json digests in these tarballs are now the only surviving record of what those two datasets contained.

Files

index.jsonl                        500 rows, one per trial — the viewer renders this
runs/<run_id>/config.json          the Harbor job config as submitted
runs/<run_id>/result.json          Harbor's job-level result and per-eval stats
runs/<run_id>/lock.json            resolved per-trial task digests and agent kwargs
runs/<run_id>/logs.tar.gz          the full raw trial dirs (transcripts, verifier output)

Each trial dir inside a tarball holds agent/pi.txt (the event log), agent/pi/sessions/*.jsonl (the session record), verifier/find_answer.txt (the graded artifact), verifier/find_metrics.txt, verifier/reward.txt, verifier/metric.err (non-empty when the scorer crashed), and result.json / lock.json.

`index.jsonl` schema

Provenance: run_id, arm, code_commit, code_commit_state, agent_class, pi_version, model, thinking, reasoning_effort_on_wire, tools, step_limit, submit_grace_steps. Dataset: dataset_source, dataset_kind, task_name, task_digest, task_checksum. Outcome: trial_name, status (scored | error), exception_type, reward, f1_score, precision, recall, hit_file_rate, hit_region_rate, n_predicted_files. Protocol diagnostics: n_turns, capped, wrote_answer_file, scorer_crashed, n_input_tokens, n_output_tokens, started_at, finished_at.

reward is f1_score for find tasks; both are kept because a crashed scorer writes neither, and the pair distinguishes "scored 0" from "no score".

Three caveats before you quote a number

Harbor's job-level mean divides by all 100 trials while labelling the block n_trials: 98. The f1 column in the table above is the mean over scored trials, which is result.json's mean × 100/98. Always report the error count next to the mean. Recompute from index.jsonl rather than reading result.json's mean directly.
Every run lost the same 2 trials, both instance_protonmail__webclients__*, both NetworkConnectionError — an nvm install exiting 127 during agent setup. It is an egress failure on the run host, not a task or model property, and it is identical across all five runs, so it does not bias any comparison.
pi drifted 0.84.3 → 0.84.4 mid-series, because the agent is installed with npm install -g @earendil-works/pi-coding-agent@latest. The A/B pair (09-06-36 vs 03-24-51) is safe — both 0.84.3. Comparing 03-24-51 (0.84.3) to 06-18-35 (0.84.4) crosses that boundary; the harness version is an uncontrolled variable in that one comparison.

Access and licensing

Gated: access is granted manually, for research and evaluation. The logs quote verbatim source from the upstream repositories under evaluation — django, sympy, matplotlib, scikit-learn, tutanota, element-web and ~40 others — each under its own license, which is why this is marked license: other rather than a single permissive license. The Harbor configs, index.jsonl and this card are ours; the quoted source excerpts inside the tarballs are not.

Reproducing this bundle

cd swe-explore-fix
python3 scripts/build_run_log_dataset.py --out ~/hf_upload/swe-explore-find-dev100-runs

The builder asserts its own per-trial walk against Harbor's n_trials / n_errors for every run, so a silently truncated log tree fails the build rather than shipping a short index.

Downloads last month: 20

Total file size:

39.2 MB

You need to agree to share your Hugging Face information to access this dataset