Datasets:
shard int64 0 11 | row int64 0 6.67k | instance_id stringlengths 10 57 | model_name stringclasses 3
values | target bool 2
classes | exit_status stringclasses 9
values | category stringclasses 5
values | n_steps int64 0 408 | n_product_files int64 0 10k | touched_tests bool 2
classes | last_product_edit float64 0 317 ⌀ | runs_after_edit int64 0 268 | tests_after_edit int64 0 110 | last_outcome stringclasses 3
values | claims_success bool 2
classes |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
0 | 0 | AnalogJ__lexicon-336 | swe-agent-llama-70b | false | submitted (exit_context) | unverified | 46 | 1 | false | 45 | 0 | 0 | null | false |
0 | 1 | AnalogJ__lexicon-336 | swe-agent-llama-70b | false | submitted (exit_context) | unverified | 25 | 1 | false | 7 | 0 | 0 | null | false |
0 | 2 | AnalogJ__lexicon-336 | swe-agent-llama-70b | false | submitted (exit_context) | unverified | 29 | 1 | false | 28 | 0 | 0 | null | false |
0 | 3 | AnalogJ__lexicon-336 | swe-agent-llama-70b | false | submitted | ran_passing | 16 | 1 | false | 10 | 1 | 0 | unknown | true |
0 | 4 | AnalogJ__lexicon-336 | swe-agent-llama-70b | false | submitted | ran_passing | 20 | 1 | false | 16 | 1 | 0 | unknown | false |
0 | 5 | AnalogJ__lexicon-336 | swe-agent-llama-70b | false | submitted | ran_failing | 9 | 1 | false | 6 | 1 | 0 | fail | false |
0 | 6 | AnalogJ__lexicon-336 | swe-agent-llama-70b | false | submitted (exit_context) | unverified | 155 | 1 | false | 11 | 0 | 0 | null | false |
0 | 7 | AnalogJ__lexicon-336 | swe-agent-llama-70b | false | submitted | ran_passing | 14 | 1 | false | 3 | 3 | 0 | unknown | true |
0 | 8 | AnalogJ__lexicon-336 | swe-agent-llama-70b | false | submitted | unverified | 11 | 1 | false | 9 | 0 | 0 | null | false |
0 | 9 | AnalogJ__lexicon-336 | swe-agent-llama-70b | false | submitted | ran_passing | 16 | 1 | false | 12 | 1 | 0 | pass | false |
0 | 10 | AnalogJ__lexicon-336 | swe-agent-llama-70b | false | submitted (exit_context) | no_product_change | 41 | 0 | true | null | 0 | 0 | null | false |
0 | 11 | Azure__azure-functions-python-worker-890 | swe-agent-llama-70b | false | submitted | unverified | 8 | 1 | false | 6 | 0 | 0 | null | false |
0 | 12 | Azure__azure-functions-python-worker-890 | swe-agent-llama-70b | false | submitted | ran_passing | 10 | 1 | false | 4 | 1 | 0 | unknown | false |
0 | 13 | Azure__azure-functions-python-worker-890 | swe-agent-llama-70b | false | submitted | unverified | 6 | 1 | false | 4 | 0 | 0 | null | true |
0 | 14 | Azure__azure-functions-python-worker-890 | swe-agent-llama-70b | false | submitted | unverified | 10 | 1 | false | 8 | 0 | 0 | null | false |
0 | 15 | Azure__azure-functions-python-worker-890 | swe-agent-llama-70b | false | submitted | unverified | 11 | 1 | false | 9 | 0 | 0 | null | true |
0 | 16 | Azure__azure-functions-python-worker-890 | swe-agent-llama-70b | false | submitted | unverified | 14 | 1 | false | 12 | 0 | 0 | null | false |
0 | 17 | Azure__azure-functions-python-worker-890 | swe-agent-llama-70b | false | submitted | no_product_change | 15 | 0 | false | null | 0 | 0 | null | true |
0 | 18 | Azure__azure-functions-python-worker-890 | swe-agent-llama-70b | false | submitted | unverified | 32 | 1 | false | 21 | 0 | 0 | null | false |
0 | 19 | Azure__azure-functions-python-worker-890 | swe-agent-llama-70b | false | submitted | unverified | 8 | 1 | false | 6 | 0 | 0 | null | false |
0 | 20 | Azure__azure-functions-python-worker-890 | swe-agent-llama-70b | false | submitted | unverified | 14 | 1 | false | 10 | 0 | 0 | null | true |
0 | 21 | Azure__azure-functions-python-worker-890 | swe-agent-llama-70b | false | submitted | ran_passing | 17 | 1 | false | 14 | 1 | 1 | pass | true |
0 | 22 | Azure__azure-functions-python-worker-890 | swe-agent-llama-70b | false | submitted | no_product_change | 11 | 0 | true | null | 0 | 0 | null | false |
0 | 23 | Azure__azure-functions-python-worker-890 | swe-agent-llama-70b | false | submitted | no_product_change | 16 | 0 | false | null | 0 | 0 | null | false |
0 | 24 | Azure__azure-functions-python-worker-890 | swe-agent-llama-70b | false | submitted | no_product_change | 12 | 0 | false | null | 0 | 0 | null | false |
0 | 25 | Azure__azure-functions-python-worker-890 | swe-agent-llama-70b | false | submitted | ran_passing | 12 | 1 | false | 4 | 2 | 0 | unknown | true |
0 | 26 | Melevir__cognitive_complexity-15 | swe-agent-llama-70b | true | submitted | ran_passing | 21 | 1 | true | 11 | 2 | 2 | pass | true |
0 | 27 | Melevir__cognitive_complexity-15 | swe-agent-llama-70b | false | submitted | ran_passing | 17 | 1 | false | 14 | 1 | 1 | pass | true |
0 | 28 | Melevir__cognitive_complexity-15 | swe-agent-llama-70b | false | submitted | no_product_change | 17 | 0 | false | null | 0 | 0 | null | true |
0 | 29 | Melevir__cognitive_complexity-15 | swe-agent-llama-8b | false | submitted | no_product_change | 10 | 0 | true | null | 0 | 0 | null | false |
0 | 30 | Melevir__cognitive_complexity-15 | swe-agent-llama-70b | false | submitted (exit_context) | ran_passing | 17 | 1 | false | 15 | 1 | 1 | unknown | false |
0 | 31 | Melevir__cognitive_complexity-15 | swe-agent-llama-8b | true | submitted | ran_failing | 15 | 1 | false | 10 | 2 | 2 | fail | true |
0 | 32 | Melevir__cognitive_complexity-15 | swe-agent-llama-70b | true | submitted (exit_context) | ran_passing | 40 | 1 | true | 13 | 8 | 0 | unknown | false |
0 | 33 | Melevir__cognitive_complexity-15 | swe-agent-llama-70b | false | submitted | no_product_change | 7 | 0 | true | null | 0 | 0 | null | false |
0 | 34 | Melevir__cognitive_complexity-15 | swe-agent-llama-70b | true | submitted | ran_passing | 17 | 1 | false | 6 | 2 | 0 | unknown | false |
0 | 35 | Melevir__cognitive_complexity-15 | swe-agent-llama-70b | true | submitted | ran_passing | 18 | 1 | true | 6 | 3 | 3 | pass | true |
0 | 36 | Melevir__cognitive_complexity-15 | swe-agent-llama-70b | true | submitted | ran_passing | 19 | 1 | true | 11 | 2 | 2 | pass | false |
0 | 37 | Melevir__cognitive_complexity-15 | swe-agent-llama-70b | true | submitted | ran_failing | 28 | 1 | true | 12 | 4 | 4 | fail | false |
0 | 38 | Melevir__cognitive_complexity-15 | swe-agent-llama-70b | false | submitted | ran_passing | 14 | 1 | false | 7 | 1 | 1 | pass | true |
0 | 39 | Melevir__cognitive_complexity-15 | swe-agent-llama-70b | false | submitted | no_product_change | 20 | 0 | false | null | 0 | 0 | null | true |
0 | 40 | Melevir__cognitive_complexity-15 | swe-agent-llama-70b | true | submitted | ran_failing | 12 | 1 | false | 9 | 1 | 1 | fail | true |
0 | 41 | Melevir__cognitive_complexity-15 | swe-agent-llama-70b | false | submitted | ran_passing | 19 | 1 | false | 15 | 1 | 0 | unknown | true |
0 | 42 | Melevir__cognitive_complexity-15 | swe-agent-llama-70b | false | submitted | ran_passing | 18 | 1 | false | 14 | 2 | 2 | pass | true |
0 | 43 | Melevir__cognitive_complexity-15 | swe-agent-llama-8b | false | submitted | no_product_change | 7 | 0 | true | null | 0 | 0 | null | true |
0 | 44 | Melevir__cognitive_complexity-15 | swe-agent-llama-70b | false | submitted (exit_context) | ran_failing | 58 | 1 | false | 28 | 1 | 0 | fail | false |
0 | 45 | Melevir__cognitive_complexity-15 | swe-agent-llama-70b | false | submitted | ran_passing | 15 | 1 | false | 12 | 1 | 1 | pass | true |
0 | 46 | Melevir__cognitive_complexity-15 | swe-agent-llama-70b | false | submitted | ran_passing | 23 | 1 | true | 13 | 2 | 2 | pass | true |
0 | 47 | Melevir__cognitive_complexity-15 | swe-agent-llama-70b | true | submitted | ran_passing | 15 | 1 | false | 11 | 1 | 0 | unknown | true |
0 | 48 | Melevir__cognitive_complexity-15 | swe-agent-llama-70b | true | submitted | unverified | 24 | 1 | true | 22 | 0 | 0 | null | true |
0 | 49 | Melevir__cognitive_complexity-15 | swe-agent-llama-70b | true | submitted (exit_context) | ran_passing | 18 | 1 | false | 16 | 1 | 1 | unknown | false |
0 | 50 | Melevir__cognitive_complexity-15 | swe-agent-llama-70b | false | submitted (exit_context) | unverified | 31 | 1 | false | 30 | 0 | 0 | null | false |
0 | 51 | Melevir__cognitive_complexity-15 | swe-agent-llama-8b | false | submitted (exit_context) | ran_failing | 44 | 1 | true | 13 | 1 | 1 | fail | false |
0 | 52 | Melevir__cognitive_complexity-15 | swe-agent-llama-70b | true | submitted | ran_failing | 15 | 1 | false | 10 | 1 | 1 | fail | false |
0 | 53 | Melevir__cognitive_complexity-15 | swe-agent-llama-70b | false | submitted | no_product_change | 7 | 0 | true | null | 0 | 0 | null | true |
0 | 54 | Melevir__cognitive_complexity-15 | swe-agent-llama-70b | false | submitted | ran_passing | 11 | 1 | false | 8 | 1 | 1 | pass | true |
0 | 55 | Melevir__cognitive_complexity-15 | swe-agent-llama-70b | false | submitted | unverified | 10 | 1 | false | 8 | 0 | 0 | null | false |
0 | 56 | Melevir__cognitive_complexity-15 | swe-agent-llama-70b | false | submitted (exit_context) | no_product_change | 68 | 0 | false | null | 0 | 0 | null | false |
0 | 57 | Melevir__cognitive_complexity-15 | swe-agent-llama-70b | false | submitted | ran_passing | 13 | 1 | false | 10 | 1 | 1 | pass | true |
0 | 58 | Melevir__cognitive_complexity-15 | swe-agent-llama-70b | true | submitted | ran_passing | 18 | 1 | false | 6 | 3 | 0 | unknown | false |
0 | 59 | Melevir__cognitive_complexity-15 | swe-agent-llama-70b | true | submitted | ran_passing | 22 | 2 | false | 18 | 1 | 0 | unknown | true |
0 | 60 | Melevir__cognitive_complexity-15 | swe-agent-llama-70b | false | submitted | ran_passing | 21 | 1 | false | 10 | 3 | 0 | pass | true |
0 | 61 | Melevir__cognitive_complexity-15 | swe-agent-llama-70b | false | submitted | ran_passing | 13 | 1 | false | 9 | 1 | 1 | pass | true |
0 | 62 | Melevir__cognitive_complexity-15 | swe-agent-llama-70b | true | submitted | ran_failing | 23 | 1 | false | 11 | 4 | 4 | fail | true |
0 | 63 | Melevir__cognitive_complexity-15 | swe-agent-llama-70b | false | submitted (exit_context) | no_product_change | 35 | 0 | false | null | 0 | 0 | null | false |
0 | 64 | Melevir__cognitive_complexity-15 | swe-agent-llama-70b | true | submitted | ran_passing | 34 | 4 | false | 30 | 1 | 0 | unknown | false |
0 | 65 | Melevir__cognitive_complexity-15 | swe-agent-llama-8b | false | submitted | ran_failing | 12 | 1 | false | 9 | 1 | 0 | fail | false |
0 | 66 | Melevir__cognitive_complexity-15 | swe-agent-llama-70b | false | submitted | ran_passing | 12 | 1 | false | 9 | 1 | 1 | pass | true |
0 | 67 | Melevir__cognitive_complexity-15 | swe-agent-llama-70b | false | submitted | unverified | 16 | 1 | false | 14 | 0 | 0 | null | true |
0 | 68 | Melevir__cognitive_complexity-15 | swe-agent-llama-70b | false | submitted | ran_failing | 28 | 1 | false | 24 | 1 | 0 | fail | true |
0 | 69 | Melevir__cognitive_complexity-15 | swe-agent-llama-70b | true | submitted | ran_passing | 19 | 1 | true | 10 | 2 | 2 | pass | true |
0 | 70 | Melevir__cognitive_complexity-15 | swe-agent-llama-70b | false | submitted | no_product_change | 14 | 0 | true | null | 0 | 0 | null | true |
0 | 71 | PyCQA__pyflakes-761 | swe-agent-llama-70b | false | submitted (exit_context) | unverified | 44 | 3 | false | 43 | 0 | 0 | null | false |
0 | 72 | PyCQA__pyflakes-761 | swe-agent-llama-70b | false | submitted (exit_context) | unverified | 39 | 4 | true | 38 | 0 | 0 | null | false |
0 | 73 | PyCQA__pyflakes-761 | swe-agent-llama-70b | false | submitted (exit_context) | unverified | 38 | 2 | false | 3 | 0 | 0 | null | false |
0 | 74 | PyCQA__pyflakes-761 | swe-agent-llama-70b | false | submitted (exit_context) | ran_failing | 47 | 3 | true | 37 | 1 | 0 | fail | false |
0 | 75 | PyCQA__pyflakes-761 | swe-agent-llama-70b | false | submitted (exit_context) | unverified | 83 | 2 | false | 4 | 0 | 0 | null | false |
0 | 76 | PyCQA__pyflakes-761 | swe-agent-llama-70b | false | submitted | ran_passing | 23 | 3 | false | 19 | 1 | 0 | unknown | false |
0 | 77 | PyCQA__pyflakes-761 | swe-agent-llama-70b | false | submitted | ran_passing | 38 | 1 | false | 21 | 12 | 0 | unknown | false |
0 | 78 | PyCQA__pyflakes-761 | swe-agent-llama-70b | false | submitted (exit_context) | unverified | 50 | 2 | false | 4 | 0 | 0 | null | false |
0 | 79 | PyCQA__pyflakes-761 | swe-agent-llama-70b | false | submitted | unverified | 26 | 1 | false | 21 | 0 | 0 | null | true |
0 | 80 | PyCQA__pyflakes-761 | swe-agent-llama-70b | false | submitted (exit_context) | ran_passing | 57 | 3 | false | 54 | 1 | 0 | unknown | false |
0 | 81 | PyCQA__pyflakes-761 | swe-agent-llama-70b | false | submitted | unverified | 28 | 1 | false | 24 | 0 | 0 | null | true |
0 | 82 | PyCQA__pyflakes-761 | swe-agent-llama-70b | false | submitted (exit_context) | unverified | 40 | 2 | false | 3 | 0 | 0 | null | false |
0 | 83 | RDFLib__rdflib-1130 | swe-agent-llama-70b | false | submitted | unverified | 4 | 1 | false | 2 | 0 | 0 | null | false |
0 | 84 | RDFLib__rdflib-1130 | swe-agent-llama-70b | false | submitted | unverified | 4 | 1 | false | 2 | 0 | 0 | null | false |
0 | 85 | RDFLib__rdflib-1130 | swe-agent-llama-70b | false | submitted | unverified | 4 | 1 | false | 2 | 0 | 0 | null | false |
0 | 86 | RDFLib__rdflib-1130 | swe-agent-llama-70b | false | submitted | unverified | 5 | 1 | false | 3 | 0 | 0 | null | false |
0 | 87 | RDFLib__rdflib-1130 | swe-agent-llama-70b | false | submitted | unverified | 3 | 1 | false | 1 | 0 | 0 | null | false |
0 | 88 | ReproNim__reproman-518 | swe-agent-llama-70b | false | submitted | unverified | 12 | 1 | false | 10 | 0 | 0 | null | false |
0 | 89 | ReproNim__reproman-518 | swe-agent-llama-70b | false | submitted (exit_context) | ran_passing | 69 | 1 | false | 8 | 20 | 0 | pass | false |
0 | 90 | ReproNim__reproman-518 | swe-agent-llama-70b | false | submitted | ran_passing | 29 | 2 | false | 24 | 2 | 0 | unknown | false |
0 | 91 | ReproNim__reproman-518 | swe-agent-llama-70b | false | exit_context | no_product_change | 5 | 0 | false | null | 0 | 0 | null | false |
0 | 92 | ReproNim__reproman-518 | swe-agent-llama-70b | false | submitted | no_product_change | 25 | 0 | true | null | 0 | 0 | null | false |
0 | 93 | ReproNim__reproman-518 | swe-agent-llama-70b | false | exit_context | no_product_change | 39 | 0 | false | null | 0 | 0 | null | false |
0 | 94 | ReproNim__reproman-518 | swe-agent-llama-70b | false | submitted | unverified | 13 | 1 | false | 11 | 0 | 0 | null | false |
0 | 95 | ShawHahnLab__igseq-2 | swe-agent-llama-70b | false | submitted | ran_failing | 32 | 1 | true | 24 | 6 | 0 | fail | false |
0 | 96 | ShawHahnLab__igseq-2 | swe-agent-llama-70b | false | submitted (exit_context) | ran_failing | 32 | 1 | true | 29 | 1 | 0 | fail | false |
0 | 97 | ShawHahnLab__igseq-2 | swe-agent-llama-70b | false | submitted (exit_context) | ran_passing | 56 | 1 | true | 50 | 1 | 0 | unknown | false |
0 | 98 | ShawHahnLab__igseq-2 | swe-agent-llama-70b | false | submitted | ran_passing | 16 | 1 | false | 4 | 4 | 0 | unknown | false |
0 | 99 | ShawHahnLab__igseq-2 | swe-agent-llama-70b | false | submitted | ran_passing | 25 | 4 | true | 5 | 4 | 0 | unknown | true |
done-claims
In 18,640 of 47,293 submitted SWE-agent runs that changed product code (39.4%), the agent ran no test or script after its last edit. Those submissions resolved the issue 18.2% of the time, against 34.0% for submissions whose last run after the edit showed no failure signal; comparing attempts at the same issue, the Mantel-Haenszel odds ratio is 0.68 (95% CI 0.63 to 0.74, 598 issues). A rerun with Claude Code driven by gpt-oss-120b on SWE-bench Verified, with its hypothesis and decision rule set in a plan dated before it started, gives 0.55 (95% CI 0.37 to 0.82, 73 issues) over the 147 issues where the agent skipped verification at least once. An earlier run with the same agent and model was inconclusive: 1.04 (0.42 to 2.55, 22 issues).
This dataset holds the measurements behind those numbers, one row per trajectory or trial, with its category, run
counts and outcome. It contains no trajectory or transcript text. The classifier, the dated plan and every
number with its source are at github.com/speker/done-claims
(projects/done-claims/RESULTS.md and PLAN.md). The write-up is the article
Coding agents that submit without running a test or script
(Türkçe).
Files and columns
| Config | Rows | One row is |
|---|---|---|
phase1_swe_agent |
80,036 | a trajectory of nebius/SWE-agent-trajectories (revision 68195a1); the analysis keeps the 51,087 whose exit status is submitted |
phase2a_claude_code |
1,500 | a Claude Code trial on SWE-bench Verified in phase 2a, reported as a description only |
phase2b_stage1 |
1,500 | a trial of stage 1: 500 issues, 3 attempts each |
phase2b_stage2 |
1,176 | a trial of stage 2: the 147 issues with an unverified stage-1 attempt, 8 more attempts each |
results/ holds the summaries, the phase 2b analysis (results/phase2b/analysis.json) and every audit file, in the
same layout as the GitHub repository. The columns:
category:unverified(no test or script ran after the last product edit),ran_failing(the last run after it failed),ran_passing(the last run after it showed no failure signal),no_product_change; phase 1 also hasundetermined(product files changed, but no edit to them is visible). The full definitions of a product file, an edit, a run and a failure are in RESULTS.md.target: whether the issue was resolved. Phase 1 takes the dataset's label (the final patch passed the tests of the linked pull request); phase 2 sets it when the SWE-bench verifier's reward is 1.n_steps,n_product_files,touched_tests(a patched or edited file is a test file),last_product_edit(step index),runs_after_edit,tests_after_edit,last_outcome(pass,fail,unknown, empty when no test or script ran),claims_success(the closing message asserts success without a hedge or a negation, by the lexicon indoneclaims/rules.py).- Phase 1 only:
shardandrow(position in the source Parquet shards),instance_id,model_name(the SWE-agent action model),exit_status. - Phase 2 only:
trial(<instance_id>__s1-NN,s2-NNin stage 2),status(completedor the exception type),reward,remote_fetches(shell commands that reached a URL or a git remote),input_tokens,output_tokens,agent_seconds.
How the categories were checked
Each phase was audited on two random samples of 40 rows (10 per category in phase 2). In phase 2b, two auditor agents built on
different model families read each of the 80 stage-1 rows separately and the agent that wrote the classifier re-read the
rows they disputed: 78 are right, one is wrong (a file under django/test/, which ships with Django, was taken for a
test) and one is ambiguous. Instead of a human re-check, a jury of four language models from different families
re-read all 80 rows without seeing those verdicts. All four agreed with the final verdict on 79 rows, the wrong
one included; on the ambiguous row two called the classifier right, one wrong and one ambiguous. No human has
re-checked the rows, so these are consistency checks by models, not independent validation. Fixing the
django/test/ rule moves 3 of the 2,540 completed phase 2b trials and leaves the H1 estimate unchanged.
Limitations
- The comparison is observational. Agents that verify may differ in ways that stratifying by issue does not capture.
- Each phase covers one harness and one model family: SWE-agent with three Llama-based action models from 2024, and Claude Code with gpt-oss-120b.
- Categories are computed by rules. A script that prints a wrong answer without raising counts as
ran_passing. - The 147 phase 2b issues are those where the agent skipped verification at least once in stage 1, so the pooled estimate describes those issues.
- The
testandtestingsubpackage error found in the phase 2b audit also applies to phases 1 and 2a. With the fixed rule, 18 of the 51,087 phase 1 submissions change category and the phase 1 odds ratio stays at 0.68 (0.63 to 0.74); in phase 2a, 2 of the 1,469 completed trials change and it becomes 0.94 (0.39 to 2.25, 23 issues). The numbers on this card keep the frozen rule.
License, sources and citation
CC BY 4.0. Phase 1 is derived from nebius/SWE-agent-trajectories (CC BY 4.0); its card asks users of the model outputs to follow the Llama 3.1 license, and no model output is included here. Phase 2 uses SWE-bench Verified instance ids and verifier results; that dataset's card declares no license, and none of its text is included.
@misc{peker2026doneclaims,
title = {done-claims: coding agents that submit without running a test or script after their last edit},
author = {Peker, Serkan},
year = {2026},
url = {https://huggingface.co/datasets/serkanpeker/done-claims}
}
- Downloads last month
- 314