Oct 16 21 days ago 10:22 Simplify SETA v2 dataset summary 10:19 Clarify SETA v2 task counts and reused environments 10:16 Clarify archived task curation rewards 10:10 Merge SETA recovery into task selection 10:00 Add completed SETA pass16 recovery evidence Adds the completed SETA pass@16 recovery tranche under `seta_v2/runs/2026-10-15-pass16-recovery/`. It contains 22,263 scheduled trial records: 13,175 Instruct and 9,088 Thinking, with saved trajectories and raw results.
The audit preserves 764 final attempts without verifier rewards and their original error messages: 524 harness failures, 63 supported agent failures, 32 verifier timeouts and 145 unresolved errors. The `reward` and `verifier_reward` fields remain null when the verifier produced no reward.
This is the remaining-attempt recovery tranche, not a standalone pass@16 estimate. It overlaps shards already included in the frozen SETA v2 partial snapshot from 2026-10-15 20:57 UTC; that snapshot remains reproducible and unchanged. 09:52 Use consistent TermiGen cached_images folder and pinned task references 09:51 Place TermiGen cached images in dataset folder (3201-3348) 09:51 Place TermiGen cached images in dataset folder (2801-3200) 09:50 Place TermiGen cached images in dataset folder (2401-2800) 09:50 Place TermiGen cached images in dataset folder (2001-2400) 09:50 Place TermiGen cached images in dataset folder (1601-2000) 09:50 Place TermiGen cached images in dataset folder (1201-1600) 09:50 Place TermiGen cached images in dataset folder (801-1200) 09:49 Place TermiGen cached images in dataset folder (401-800) 09:49 Place TermiGen cached images in dataset folder (1-400) 09:47 Document published TermiGen cached-image coverage 09:47 Add TermiGen cached images and v2 teacher trajectories (#20 ) FWeindel +1: llincke - Add TermiGen v1 cached image sidecar (c785ec5df519b74f49c710b95b814ec04a0a2b25)
- Add files using upload-large-folder tool (d387653408cd4846eb4aba4fbefdadc5db971595)
- Add files using upload-large-folder tool (db7541e5f7ca7251fd325a1fc8e6319ff602af66)
- Add files using upload-large-folder tool (2a4ba33f0c4bfde1ab177bd6669a2d2b9dff0bd2)
- Add files using upload-large-folder tool (8d0152c87f91e79f64eedc1f07f708f25da53cea)
- Add files using upload-large-folder tool (e7e63384743916a2d174462633b949e939d8f0ff)
- Add files using upload-large-folder tool (e0338464024f0657114a02b259dfa42407422bd6)
- Add files using upload-large-folder tool (39bed9b53ddd225a616c3bebf8038ef441f58ea3)
- Add files using upload-large-folder tool (aaba98126969be615bfb1c13475d7e02ce527f4b)
- Add files using upload-large-folder tool (cd504e0db479ccb05cfb5efcc9a274c5a1898d2d)
- Add files using upload-large-folder tool (02db7e53cc7dfe654a7aef50252b81e90ab2b5c5)
- Add files using upload-large-folder tool (6ec77c6b7df8b9… 09:10 Publish TermiGen v2 teacher traces from PR #20 Copy the 70 trace shards and coverage manifest from 0d280d415129ded45c5a798c44ced190a485dea5, preserving their contents, and link them from the v2 README. 09:09 Rename v2 selection assets folders to metadata Move eight selection manifests to metadata and update reproduction scripts and README references. 07:52 Clarify CalibForge static-check findings and missing coverage Describe the instruction suffix as appended, report 275 kept tasks with remaining absolute-path warnings, and disclose that file-reference passes skipped comparisons because no reference solutions exist. 07:49 Simplify README wording and put rule-based checks last Use kept instead of retained in README prose. Order per-stage breakdowns as build/setup, no-op, reference solution, then rule-based checks. Preserve task counts, validation findings and schema field names. Oct 15 22 days ago 22:10 Keep CalibForge pilot review results in GitHub 22:00 Summarize v1 and v2 task totals and source proportions 21:51 Keep ScaleSWE patch inputs under assets 21:48 Use consistent short dataset names with v1 and v2 folders 21:32 Publish EndlessTerminals v2 with pass@4 selection and complete traces 21:29 Publish partial ScaleSWE and SWE-Lego v2 with snapshot traces 21:11 Publish partial CalibForge v2 and verified SETA and CalibForge traces 21:01 Publish SETA v2 using verified partial teacher results 20:55 Publish verified compact successful and failed v2 teacher traces 20:51 Publish TermiGen v2 and reuse v1 excluded task payloads 20:43 Publish crosscodeeval_v2 filtered by both models pass@16 results Retain 863 of 6710 v1 tasks solved by Qwen3-30B Instruct or Thinking; archive 5847 newly unsolved tasks alongside 1053 prior exclusions. Preserve all task payloads, reuse pinned v1 images, and expose per-model success counts and selection examples. 20:40 Publish busginpy_v2 filtered by both models pass@16 results Retain 65 of 489 v1 tasks solved by Qwen3-30B Instruct or Thinking; archive 424 newly unsolved tasks alongside 11 prior exclusions. Preserve all task payloads, reuse pinned v1 images, and expose per-model success counts and selection examples. 20:29 Add CalibForge pilot judge outcomes and explanations 171 completed Opus 5.5 reviews from a 1,000-task sample. Unfinished or errored reviews are excluded; these findings are model judgments. 20:23 Keep dataset runtime summaries concise and show only available timings 20:19 Report both mean and median validation times where task records support them Derive retained-task ScaleSWE and TermiGen statistics from saved per-task reports; preserve the distinction between validation and image-build time. 20:16 Simplify runtime versions and per-check timing across all dataset cards Link the pinned Marin Harbor implementation and local runtime patches. Separate measured means from recorded medians, preserve sample sizes and missing timing data, and remove standalone historical-evidence links and unrepresentative repair timings. 20:12 Use descriptive validation labels in all dataset READMEs Remove runner stage numbers, preserve counts and evidence, and clarify that ScaleSWE uses upstream Docker Hub images with optional Apptainer caches. 19:50 Clarify historical TermiGen static exclusion reasons 19:45 Restore reviewed repairs and concise release evidence for camel-ai__SETA-Env_environment-reliable 19:45 Restore reviewed repairs and concise release evidence for EndlessTerminals_environment-reliable 19:44 Restore reviewed repairs and concise release evidence for Calibforge_environment-reliable 19:44 Clarify historical TermiGen static exclusion reasons 19:43 Add verified pristine images for restored camel-ai__SETA-Env_environment-reliable tasks 19:42 Add verified pristine images for restored Calibforge_environment-reliable tasks 19:37 Restore ScaleSWE historical static repairs and concise release evidence 19:36 Restore SWE-Lego historical static repairs and concise release evidence 19:34 Publish 48 verified ScaleSWE cached Apptainer bundles 19:30 Restore BugsInPy historical static repairs with verified cached images 19:28 Document TermiGen retained repairs, static exclusions and runtime evidence 
