fineenvs_v1: environment images for 2026-10-20-fineenvs_v1-36fc78e213a3

### Reproducibility and retries

| Stage | Mode | Tasks with expectation | Matched first attempt | Matched after 1 / 2 / 3+ retries | Archived as not robust | Not-run retried | Recovered | Unresolved |
| --- | --- | --- | --- | --- | --- | --- | --- | --- |
| 1 | classifier | 0 | 0 | 0 / 0 / 0 | 0 | 0 | 0 | 0 |
| 2 | classifier | 0 | 0 | 0 / 0 / 0 | 0 | 0 | 0 | 0 |
| 3 | classifier | 0 | 0 | 0 / 0 / 0 | 0 | 0 | 0 | 0 |

Expected outcomes come from an earlier audit of the same tasks (matched by task ID). A task whose result differs from that audit is rerun up to the retry limit; a task that never matches is archived as not robust. Without an expectation only tasks with no result (crash, node failure, missing reward or missing runner evidence) are rerun, and a task still without a result after the last attempt is archived as not robust. Stages 1, 2 and 3 are covered; the retries run after the whole pipeline.

## fineenvs_v1

Keeps **5,000 tasks** and archives **0** from [FineEnvs/SmolDataEnvs-harbor-train](https://huggingface.co/datasets/FineEnvs/SmolDataEnvs-harbor-train) at commit `4719635555de1666f374d847baebb500d368e493`.


Our patcher modified **5,000 of the original 5,000 tasks**.

Automatic fixes changed **4,576 unique tasks**.

Tasks can have multiple change labels, so category counts may overlap.

| Change label | Tasks affected | Explanation |
| --- | ---: | --- |
| gold-answer-solution | 5000 | Retain one shared upstream image pinned to docker.io/savatar101/env-data-agent-train@sha256:7b3f7e22780b431ce0d43717dcef6875cde84baeee824cc87e968374d62d3470; use its existing math-verify 0.9.0 and check installed dependency versions in tests/setup.sh instead of repeating a best-effort pip install. Move the original input initialization for BUCKET_PREFIX=4quant__soft-tissue-sarcoma and nonempty-input check into setup_files/setup.sh and remove the duplicate healthcheck hook. Preserve the upstream input basename mapping, environment variables, instructions, grading logic and reward.txt output format. Rename the JSON correctness score to reward, including empty/missing-answer zeros; preserve the score value and auxiliary metrics. Add solution/solve.sh, which writes the verifier's EXPECTED_ANSWER to /workdir/answer.txt, so the reference stage can check that the environment and grader accept it; it does not recompute the answer from the data. Package 3 input files (397355382 bytes) under setup_files/inputs with remote Xet identities and SHA-256 checksums in input-manifest.json. Remove the setup downloader invocation and copy the packaged files locally during setup, verifying their checksums with bounded-memory copying. No input download occurs at setup or image build time.<br>Retain one shared upstream image pinned to docker.io/savatar101/env-data-agent-train@sha256:7b3f7e22780b431ce0d43717dcef6875cde84baeee824cc87e968374d62d3470; use its existing math-verify 0.9.0 and check installed dependency versions in tests/setup.sh instead of repeating a best-effort pip install. Move the original input initialization for BUCKET_PREFIX=CooperUnion__anime-recommendations-database and nonempty-input check into setup_files/setup.sh and remove the duplicate healthcheck hook. Preserve the upstream input basename mapping, environment variables, instructions, grading logic and reward.json output format. Rename the JSON correctness score to reward, including empty/missing-answer zeros; preserve the score value and auxiliary metrics. Add solution/solve.sh, which writes the verifier's EXPECTED_ANSWER to /workdir/answer.txt, so the reference stage can check that the environment and grader accept it; it does not recompute the answer from the data. Package 2 input files (112341362 bytes) under setup_files/inputs with remote Xet identities and SHA-256 checksums in input-manifest.json. Remove the setup downloader invocation and copy the packaged files locally during setup, verifying their checksums with bounded-memory copying. No input download occurs at setup or image build time.<br>Retain one shared upstream image pinned to docker.io/savatar101/env-data-agent-train@sha256:7b3f7e22780b431ce0d43717dcef6875cde84baeee824cc87e968374d62d3470; use its existing math-verify 0.9.0 and check installed dependency versions in tests/setup.sh instead of repeating a best-effort pip install. Move the original input initialization for BUCKET_PREFIX=CooperUnion__anime-recommendations-database and nonempty-input check into setup_files/setup.sh and remove the duplicate healthcheck hook. Preserve the upstream input basename mapping, environment variables, instructions, grading logic and reward.txt output format. Rename the JSON correctness score to reward, including empty/missing-answer zeros; preserve the score value and auxiliary metrics. Add solution/solve.sh, which writes the verifier's EXPECTED_ANSWER to /workdir/answer.txt, so the reference stage can check that the environment and grader accept it; it does not recompute the answer from the data. Package 2 input files (112341362 bytes) under setup_files/inputs with remote Xet identities and SHA-256 checksums in input-manifest.json. Remove the setup downloader invocation and copy the packaged files locally during setup, verifying their checksums with bounded-memory copying. No input download occurs at setup or image build time.<br>Retain one shared upstream image pinned to docker.io/savatar101/env-data-agent-train@sha256:7b3f7e22780b431ce0d43717dcef6875cde84baeee824cc87e968374d62d3470; use its existing math-verify 0.9.0 and check installed dependency versions in tests/setup.sh instead of repeating a best-effort pip install. Move the original input initialization for BUCKET_PREFIX=CooperUnion__cardataset and nonempty-input check into setup_files/setup.sh and remove the duplicate healthcheck hook. Preserve the upstream input basename mapping, environment variables, instructions, grading logic and reward.json output format. Rename the JSON correctness score to reward, including empty/missing-answer zeros; preserve the score value and auxiliary metrics. Add solution/solve.sh, which writes the verifier's EXPECTED_ANSWER to /workdir/answer.txt, so the reference stage can check that the environment and grader accept it; it does not recompute the answer from the data. Package 1 input files (1475504 bytes) under setup_files/inputs with remote Xet identities and SHA-256 checksums in input-manifest.json. Remove the setup downloader invocation and copy the packaged files locally during setup, verifying their checksums with bounded-memory copying. No input download occurs at setup or image build time.<br>Retain one shared upstream image pinned to docker.io/savatar101/env-data-agent-train@sha256:7b3f7e22780b431ce0d43717dcef6875cde84baeee824cc87e968374d62d3470; use its existing math-verify 0.9.0 and check installed dependency versions in tests/setup.sh instead of repeating a best-effort pip install. Move the original input initialization for BUCKET_PREFIX=CooperUnion__cardataset and nonempty-input check into setup_files/setup.sh and remove the duplicate healthcheck hook. Preserve the upstream input basename mapping, environment variables, instructions, grading logic and reward.txt output format. Rename the JSON correctness score to reward, including empty/missing-answer zeros; preserve the score value and auxiliary metrics. Add solution/solve.sh, which writes the verifier's EXPECTED_ANSWER to /workdir/answer.txt, so the reference stage can check that the environment and grader accept it; it does not recompute the answer from the data. Package 1 input files (1475504 bytes) under setup_files/inputs with remote Xet identities and SHA-256 checksums in input-manifest.json. Remove the setup downloader invocation and copy the packaged files locally during setup, verifying their checksums with bounded-memory copying. No input download occurs at setup or image build time.<br>Retain one shared upstream image pinned to docker.io/savatar101/env-data-agent-train@sha256:7b3f7e22780b431ce0d43717dcef6875cde84baeee824cc87e968374d62d3470; use its existing math-verify 0.9.0 and check installed dependency versions in tests/setup.sh instead of repeating a best-effort pip install. Move the original input initialization for BUCKET_PREFIX=Cornell-University__movie-dialog-corpus and nonempty-input check into setup_files/setup.sh and remove the duplicate healthcheck hook. Preserve the upstream input basename mapping, environment variables, instructions, grading logic and reward.json output format. Rename the JSON correctness score to reward, including empty/missing-answer zeros; preserve the score value and auxiliary metrics. Add solution/solve.sh, which writes the verifier's EXPECTED_ANSWER to /workdir/answer.txt, so the reference stage can check that the environment and grader accept it; it does not recompute the answer from the data. Package 6 input files (30116727 bytes) under setup_files/inputs with remote Xet identities and SHA-256 checksums in input-manifest.json. Remove the setup downloader invocation and copy the packaged files locally during setup, verifying their checksums with bounded-memory copying. No input download occurs at setup or image build time.<br>Retain one shared upstream image pinned to docker.io/savatar101/env-data-agent-train@sha256:7b3f7e22780b431ce0d43717dcef6875cde84baeee824cc87e968374d62d3470; use its existing math-verify 0.9.0 and check installed dependency versions in tests/setup.sh instead of repeating a best-effort pip install. Move the original input initialization for BUCKET_PREFIX=PromptCloudHQ__amazon-reviews-unlocked-mobile-phones and nonempty-input check into setup_files/setup.sh and remove the duplicate healthcheck hook. Preserve the upstream input basename mapping, environment variables, instructions, grading logic and reward.txt output format. Rename the JSON correctness score to reward, including empty/missing-answer zeros; preserve the score value and auxiliary metrics. Add solution/solve.sh, which writes the verifier's EXPECTED_ANSWER to /workdir/answer.txt, so the reference stage can check that the environment and grader accept it; it does not recompute the answer

.gitattributes

@@ -31595,3 +31595,4 @@ termigen_v1/cached_images/bundles-v1/24dd24246b8fb450cb7caf53dfdeab1591d75bcbc4b
 scaleswe_v2/metadata/selection.json filter=lfs diff=lfs merge=lfs -text
 seta_v2/metadata/selection.json filter=lfs diff=lfs merge=lfs -text
 mimo_v1/validation.json filter=lfs diff=lfs merge=lfs -text
+fineenvs_v1/cached_images/bundles-v1/87065f5bed10a0d82dd9ce14b234658f4cf8651035df6d9864bb4c15fba5c0cf/build_0000_324_324276_qa_3-1ddd4dba12a5.sif/build_0000_324_324276_qa_3-1ddd4dba12a5.sif filter=lfs diff=lfs merge=lfs -text

fineenvs_v1/cached_images/bundles-v1/87065f5bed10a0d82dd9ce14b234658f4cf8651035df6d9864bb4c15fba5c0cf/build_0000_324_324276_qa_3-1ddd4dba12a5.sif/build_0000_324_324276_qa_3-1ddd4dba12a5.sif

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e1bb07a6c496cf961781e29a3fe728d96eebd8c3594c0c08eaa1f5afca43ad17
+size 280838144