SmolDataEnvs Multi-harness RL
Collection
Try the three environments, follow the tutorial, and explore the datasets, models and historical RL/SFT results. • 21 items • Updated • 2
task_id stringlengths 20 23 | difficulty stringclasses 2
values | harness stringclasses 1
value | rollout_id stringlengths 25 25 | teacher_model stringclasses 1
value | teacher_revision stringclasses 1
value | tokenizer_revision stringclasses 1
value | harness_version stringclasses 1
value | logprobs_mode stringclasses 1
value | provenance_json stringlengths 620 675 | conversations_json stringlengths 51k 724k | artifact_path stringclasses 13
values | artifact_member stringlengths 30 33 | source_sha256 stringlengths 64 64 | attempt int32 1 3 | correctness float64 1 1 | rollout_weight float64 1 1 | turn_count int32 3 16 | retained_tokens int64 306 12.5k | training_turns listlengths 3 16 | trace_paths listlengths 3 19 |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
0119_907_119907848_qa_1 | hard | claude-code | sdee0026474d3c1a1231302e5 | Qwen/Qwen3.8-27B | 1d4bf0f2ff6012fd82039f2fa52739d0dd7c60c0 | 1d4bf0f2ff6012fd82039f2fa52739d0dd7c60c0 | 2.1.270 | processed_logprobs | "{\"task_id\":\"0119_907_119907848_qa_1\",\"difficulty\":\"hard\",\"harness\":\"claude-code\",\"atte(...TRUNCATED) | "[{\"root_id\":\"b6efdf13a974\",\"role\":\"agent\",\"n_turns\":1,\"messages\":[{\"role\":\"system\",(...TRUNCATED) | artifacts/claude-code/part-00000.tar | 0119_907_119907848_qa_1/attempt-1 | c572f8f9664db04ac84d580b85b0b125f2efa429cccec6bca2f2d468eafac67d | 1 | 1 | 1 | 5 | 2,170 | [{"prompt_token_ids":[248045,8678,198,24342,286,4879,369,716,310,830,11553,13,5044,1683,15060,1472,2(...TRUNCATED) | ["traces/claude-code/sd/sdee0026474d3c1a1231302e5-root-000.jsonl","traces/claude-code/sd/sdee0026474(...TRUNCATED) |
0107_327_107327801_qa_2 | medium | claude-code | s38cf0ec73dca5d91e348ba29 | Qwen/Qwen3.8-27B | 1d4bf0f2ff6012fd82039f2fa52739d0dd7c60c0 | 1d4bf0f2ff6012fd82039f2fa52739d0dd7c60c0 | 2.1.270 | processed_logprobs | "{\"task_id\":\"0107_327_107327801_qa_2\",\"difficulty\":\"medium\",\"harness\":\"claude-code\",\"at(...TRUNCATED) | "[{\"root_id\":\"04f21bbe64bf\",\"role\":\"agent\",\"n_turns\":1,\"messages\":[{\"role\":\"system\",(...TRUNCATED) | artifacts/claude-code/part-00000.tar | 0107_327_107327801_qa_2/attempt-1 | eca05ee9c62c42171c621a0e908293b48ae39dda3ca1c29b664e256ca06e4497 | 1 | 1 | 1 | 5 | 905 | [{"prompt_token_ids":[248045,8678,198,24342,286,4879,369,716,310,830,11553,13,5044,1683,15060,1472,2(...TRUNCATED) | ["traces/claude-code/s3/s38cf0ec73dca5d91e348ba29-root-000.jsonl","traces/claude-code/s3/s38cf0ec73d(...TRUNCATED) |
0087_372_87372002_qa_1 | hard | claude-code | s6b0bb3c96fb6b911485e74c3 | Qwen/Qwen3.8-27B | 1d4bf0f2ff6012fd82039f2fa52739d0dd7c60c0 | 1d4bf0f2ff6012fd82039f2fa52739d0dd7c60c0 | 2.1.270 | processed_logprobs | "{\"task_id\":\"0087_372_87372002_qa_1\",\"difficulty\":\"hard\",\"harness\":\"claude-code\",\"attem(...TRUNCATED) | "[{\"root_id\":\"902d90238b50\",\"role\":\"agent\",\"n_turns\":1,\"messages\":[{\"role\":\"system\",(...TRUNCATED) | artifacts/claude-code/part-00000.tar | 0087_372_87372002_qa_1/attempt-1 | a4ca46b07160b6769a593b2cf0f9c43bf9513c63b5c5097927fe79b6a642630a | 1 | 1 | 1 | 5 | 1,122 | [{"prompt_token_ids":[248045,8678,198,24342,286,4879,369,716,310,830,11553,13,5044,1683,15060,1472,2(...TRUNCATED) | ["traces/claude-code/s6/s6b0bb3c96fb6b911485e74c3-root-000.jsonl","traces/claude-code/s6/s6b0bb3c96f(...TRUNCATED) |
0010_624_10624322_qa_4 | medium | claude-code | sb6ffdc503ad3e489ab335115 | Qwen/Qwen3.8-27B | 1d4bf0f2ff6012fd82039f2fa52739d0dd7c60c0 | 1d4bf0f2ff6012fd82039f2fa52739d0dd7c60c0 | 2.1.270 | processed_logprobs | "{\"task_id\":\"0010_624_10624322_qa_4\",\"difficulty\":\"medium\",\"harness\":\"claude-code\",\"att(...TRUNCATED) | "[{\"root_id\":\"6afb927ee104\",\"role\":\"agent\",\"n_turns\":1,\"messages\":[{\"role\":\"system\",(...TRUNCATED) | artifacts/claude-code/part-00000.tar | 0010_624_10624322_qa_4/attempt-1 | f6b1a3c465cf9f7eacd0c515ecaa87d863515ff9bac6950585f48448e32b806f | 1 | 1 | 1 | 5 | 1,170 | [{"prompt_token_ids":[248045,8678,198,24342,286,4879,369,716,310,830,11553,13,5044,1683,15060,1472,2(...TRUNCATED) | ["traces/claude-code/sb/sb6ffdc503ad3e489ab335115-root-000.jsonl","traces/claude-code/sb/sb6ffdc503a(...TRUNCATED) |
0062_391_62391096_qa_3 | hard | claude-code | s1267e3fe1e0da9e325002982 | Qwen/Qwen3.8-27B | 1d4bf0f2ff6012fd82039f2fa52739d0dd7c60c0 | 1d4bf0f2ff6012fd82039f2fa52739d0dd7c60c0 | 2.1.270 | processed_logprobs | "{\"task_id\":\"0062_391_62391096_qa_3\",\"difficulty\":\"hard\",\"harness\":\"claude-code\",\"attem(...TRUNCATED) | "[{\"root_id\":\"3c86f672b694\",\"role\":\"agent\",\"n_turns\":1,\"messages\":[{\"role\":\"system\",(...TRUNCATED) | artifacts/claude-code/part-00000.tar | 0062_391_62391096_qa_3/attempt-1 | 718613e806749da6a29134967b1fc87bcdcb992c8035ab546ca8179a2fb098ba | 1 | 1 | 1 | 4 | 687 | [{"prompt_token_ids":[248045,8678,198,24342,286,4879,369,716,310,830,11553,13,5044,1683,15060,1472,2(...TRUNCATED) | ["traces/claude-code/s1/s1267e3fe1e0da9e325002982-root-000.jsonl","traces/claude-code/s1/s1267e3fe1e(...TRUNCATED) |
0037_895_37895109_qa_2 | hard | claude-code | s341a2ad3d47152434a89fd98 | Qwen/Qwen3.8-27B | 1d4bf0f2ff6012fd82039f2fa52739d0dd7c60c0 | 1d4bf0f2ff6012fd82039f2fa52739d0dd7c60c0 | 2.1.270 | processed_logprobs | "{\"task_id\":\"0037_895_37895109_qa_2\",\"difficulty\":\"hard\",\"harness\":\"claude-code\",\"attem(...TRUNCATED) | "[{\"root_id\":\"f8bc4ab44d7d\",\"role\":\"agent\",\"n_turns\":1,\"messages\":[{\"role\":\"system\",(...TRUNCATED) | artifacts/claude-code/part-00000.tar | 0037_895_37895109_qa_2/attempt-1 | 03a4d48432fefff88615809c62ac8bb76a50ee399b87d0ad03bf41982015a746 | 1 | 1 | 1 | 6 | 2,602 | [{"prompt_token_ids":[248045,8678,198,24342,286,4879,369,716,310,830,11553,13,5044,1683,15060,1472,2(...TRUNCATED) | ["traces/claude-code/s3/s341a2ad3d47152434a89fd98-root-000.jsonl","traces/claude-code/s3/s341a2ad3d4(...TRUNCATED) |
0019_978_19978878_qa_2 | medium | claude-code | s216daeb6b04f0380f6cdc5bc | Qwen/Qwen3.8-27B | 1d4bf0f2ff6012fd82039f2fa52739d0dd7c60c0 | 1d4bf0f2ff6012fd82039f2fa52739d0dd7c60c0 | 2.1.270 | processed_logprobs | "{\"task_id\":\"0019_978_19978878_qa_2\",\"difficulty\":\"medium\",\"harness\":\"claude-code\",\"att(...TRUNCATED) | "[{\"root_id\":\"ab73d4fd444a\",\"role\":\"agent\",\"n_turns\":1,\"messages\":[{\"role\":\"system\",(...TRUNCATED) | artifacts/claude-code/part-00000.tar | 0019_978_19978878_qa_2/attempt-1 | 981dc9654c3097996e18630c3938da10ec21382a265dbb878154b5b591d0b35b | 1 | 1 | 1 | 4 | 1,071 | [{"prompt_token_ids":[248045,8678,198,24342,286,4879,369,716,310,830,11553,13,5044,1683,15060,1472,2(...TRUNCATED) | ["traces/claude-code/s2/s216daeb6b04f0380f6cdc5bc-root-000.jsonl","traces/claude-code/s2/s216daeb6b0(...TRUNCATED) |
0119_590_119590030_qa_2 | hard | claude-code | sdbea581eb497c73ee30c2c0d | Qwen/Qwen3.8-27B | 1d4bf0f2ff6012fd82039f2fa52739d0dd7c60c0 | 1d4bf0f2ff6012fd82039f2fa52739d0dd7c60c0 | 2.1.270 | processed_logprobs | "{\"task_id\":\"0119_590_119590030_qa_2\",\"difficulty\":\"hard\",\"harness\":\"claude-code\",\"atte(...TRUNCATED) | "[{\"root_id\":\"4a7488126837\",\"role\":\"agent\",\"n_turns\":1,\"messages\":[{\"role\":\"system\",(...TRUNCATED) | artifacts/claude-code/part-00000.tar | 0119_590_119590030_qa_2/attempt-2 | f42659244b6aa7d43abaca78931ae518f8f8d2a5dffc8f15fb063215f599bb0f | 2 | 1 | 1 | 9 | 8,411 | [{"prompt_token_ids":[248045,8678,198,24342,286,4879,369,716,310,830,11553,13,5044,1683,15060,1472,2(...TRUNCATED) | ["traces/claude-code/sd/sdbea581eb497c73ee30c2c0d-root-000.jsonl","traces/claude-code/sd/sdbea581eb4(...TRUNCATED) |
0100_986_100986965_qa_1 | medium | claude-code | s2e3d1bf62f400b7882a86cb1 | Qwen/Qwen3.8-27B | 1d4bf0f2ff6012fd82039f2fa52739d0dd7c60c0 | 1d4bf0f2ff6012fd82039f2fa52739d0dd7c60c0 | 2.1.270 | processed_logprobs | "{\"task_id\":\"0100_986_100986965_qa_1\",\"difficulty\":\"medium\",\"harness\":\"claude-code\",\"at(...TRUNCATED) | "[{\"root_id\":\"44e136190f46\",\"role\":\"agent\",\"n_turns\":1,\"messages\":[{\"role\":\"system\",(...TRUNCATED) | artifacts/claude-code/part-00000.tar | 0100_986_100986965_qa_1/attempt-2 | a94e0ece0e6b72600226918aee736f27cf0a0c0adb1b3d35795243bcb81010e8 | 2 | 1 | 1 | 8 | 4,215 | [{"prompt_token_ids":[248045,8678,198,24342,286,4879,369,716,310,830,11553,13,5044,1683,15060,1472,2(...TRUNCATED) | ["traces/claude-code/s2/s2e3d1bf62f400b7882a86cb1-root-000.jsonl","traces/claude-code/s2/s2e3d1bf62f(...TRUNCATED) |
0065_789_65789409_qa_2 | hard | claude-code | sc2eb04297138680c6bca5b54 | Qwen/Qwen3.8-27B | 1d4bf0f2ff6012fd82039f2fa52739d0dd7c60c0 | 1d4bf0f2ff6012fd82039f2fa52739d0dd7c60c0 | 2.1.270 | processed_logprobs | "{\"task_id\":\"0065_789_65789409_qa_2\",\"difficulty\":\"hard\",\"harness\":\"claude-code\",\"attem(...TRUNCATED) | "[{\"root_id\":\"67bf5ca5dd3b\",\"role\":\"agent\",\"n_turns\":1,\"messages\":[{\"role\":\"system\",(...TRUNCATED) | artifacts/claude-code/part-00000.tar | 0065_789_65789409_qa_2/attempt-3 | 4fa5f4daf7fb7362a08042c62ed62770e887b9714d7e376efa47615988f8e3b5 | 3 | 1 | 1 | 11 | 7,211 | [{"prompt_token_ids":[248045,8678,198,24342,286,4879,369,716,310,830,11553,13,5044,1683,15060,1472,2(...TRUNCATED) | ["traces/claude-code/sc/sc2eb04297138680c6bca5b54-root-000.jsonl","traces/claude-code/sc/sc2eb042971(...TRUNCATED) |
End of preview. Expand in Data Studio
Qwen3.8-27B Harbor Rollouts
One row is one successful rollout with nested exact training records. Collected through OpenEnv x Harbor on Daytona, with up to three attempts per task/harness and the first reward=1, TiTO-valid success retained. This is selected SFT data, not pass@1 evaluation. The fixed source set contains 400 medium and 600 hard tasks.
| Harness | Published successful rollouts | Training records | Excluded credential-bearing rollouts |
|---|---|---|---|
| opencode | 801 | 4825 | 2 |
| mini-swe-agent | 810 | 4445 | 2 |
| claude-code | 781 | 4581 | 0 |
| codex | 797 | 4078 | 0 |
Exact training data
training_turns preserves prompt_token_ids, completion_token_ids, per_token_logps (float64), and loss_mask. Logprobs align to completion tokens; masks span prompt + completion. No top-k or full-vocabulary logits were collected. These are teacher processed logprobs and teacher tokenizer IDs, not arbitrary student-token IDs. No retokenization, padding probabilities or float downcasting is applied. Credential-bearing rollouts are excluded wholesale, never text-redacted while retaining incompatible token arrays; manifests/*-excluded.json records excluded task IDs. Exact originals remain local and are not uploaded. Message/tool/metadata payloads are lossless JSON strings (prompt_json, completion_json, tools_json, metadata_json). provenance_json retains model/tokenizer revisions, sampling, runtime identity and other original rollout metadata. conversations_json retains the original conversation roots. Training records may span several model calls and must not be confused with raw request counts.
Loading
from datasets import load_dataset
ds = load_dataset("FineEnvs/qwen38-27b-harbor-rollouts", "opencode", split="train", streaming=True)
Select a harness explicitly. For training, preserve rollout grouping/weighting; flattening every training record and sampling uniformly changes rollout weighting. manifests/*-rollouts.json provides task IDs for common-task intersection and balanced sampling. All harnesses use the same task set; partition by task ID across harnesses if creating a validation split.
Traces and artifacts
trace_paths points to derived HF Session Trace Simple Format JSONL files, one per conversation root; no distinct roots are concatenated. Native session JSONLs, where available, and exact successful SFT/capture/result/contract/audit/outcome artifacts are retained in tar files addressed by artifact_path and artifact_member. Derived trace rendering is not yet verified in Data Studio; numeric Parquet data is the training source of truth. Original absolute paths in archived provenance are historical, not required to load the dataset. manifests/*-attempts.json records failed and interrupted attempts separately from SFT successes. File hashes are recorded in each harness checksum manifest.
Publication is incremental: only completed harnesses appear as configurations. The remaining subsets will be added after collection and export validation. Dataset visibility is private; no public redistribution license is asserted.
- Downloads last month
- 565