biscet commited on
Commit
8efccb2
·
verified ·
1 Parent(s): e9ae9c2

Release v1.0.1: trim text mix to ~50%, add stage-2 checkpoint and agent validation/test splits

Browse files

Removed four complete public-domain literary sources (8,285 tasks), kept the main task schema stable, published 272 anonymised stage-2 contexts as stage2/added_2026-11-13.jsonl, added 90/5/5 splits for agent trajectories and steps, and rebuilt reports, manifest and README.

README.md CHANGED
@@ -729,7 +729,7 @@ dataset_info:
729
 
730
  **DeepSeek teacher traces. Contexts for the next draft.**
731
 
732
- `41,987 task contexts`  ·  `200 agent trajectories`  ·  `294 next-action steps`  ·  `RU / EN`
733
 
734
  [Explore the data](https://huggingface.co/datasets/biscet/DFlash-DeepSeek-TraceMix/tree/main) · [Manifest](https://huggingface.co/datasets/biscet/DFlash-DeepSeek-TraceMix/blob/main/manifest.json) · [Privacy report](https://huggingface.co/datasets/biscet/DFlash-DeepSeek-TraceMix/blob/main/privacy_report.json)
735
 
@@ -741,7 +741,7 @@ dataset_info:
741
  > A bilingual mix of task prompts and DeepSeek-generated agent traces for
742
  > preparing draft-model training and studying tool use.
743
 
744
- The main corpus contains **41,987 prompts and contexts**, not DeepSeek
745
  completions. The separate agent supplements contain actual reasoning, historical
746
  tool calls, real observations and final responses distilled from the DeepSeek
747
  teacher models listed below through Qwen Code.
@@ -750,72 +750,75 @@ teacher models listed below through Qwen Code.
750
 
751
  - **Text corpus trimmed.** Four complete public-domain literary sources were
752
  removed (8,285 tasks in total) to bring the text share down to about
753
- one half of the corpus.
754
- - **New stage-2 tasks added.** 272 tasks built from two real codebases
755
  (`project_03`, `project_05_mobile`) were converted to the release schema,
756
- anonymised and placed into the matching splits.
757
- - **Resulting mix:** 52.2% text/reasoning and 47.8% software
758
- development, with 20,685 Russian and 21,302 English tasks.
 
759
  - **Splits rebalanced** proportionally to 90 / 5 / 5.
760
 
761
  ## Stage-2 checkpoint (2026-11-13)
762
 
763
- The exact tasks added in this release are also published as a standalone
764
  checkpoint, so they can be pulled out of the full dataset without filtering:
765
 
766
  [`stage2/added_2026-11-13.jsonl`](https://huggingface.co/datasets/biscet/DFlash-DeepSeek-TraceMix/blob/main/stage2/added_2026-11-13.jsonl)
767
 
768
- - **272 records**, one JSON object per line, same schema as `data/train.jsonl`.
 
 
 
 
 
769
  - Every record carries `provenance.generator_version = "1.0.1+stage2-2026-11-13"`.
770
  - `source_sha256` is computed over the anonymised message/tool context, not over
771
  the original repository file.
772
- - They are the same objects that appear in the task splits (190 train,
773
- 42 validation, 40 test), so loading both would duplicate
774
- them; use the checkpoint for extraction and `data/*.jsonl` for the complete
775
- training mix.
776
 
777
  ## Teacher models
778
 
779
  We distilled the agent traces from **DeepSeek Flash v4.1 (1017)** and
780
  **DeepSeek Pro V4 (0813)**.
781
 
782
- | Teacher model | Recorded API alias | Trajectories |
783
- |:---|:---|---:|
784
- | DeepSeek Flash v4.1 (1017) | `deepseek-flash` | 161 |
785
- | DeepSeek Pro V4 (0813) | `deepseek-v4-pro` | 39 |
786
 
 
 
787
  Trace records retain the original API aliases in their model and provenance fields.
788
 
789
  ## Repository branches
790
 
791
  | Branch | Contents |
792
  |:---|:---|
793
- | `main` | Latest complete release: task contexts, DeepSeek agent data, capture exports, manifests and reports. |
794
- | [`v1.0.1`](https://huggingface.co/datasets/biscet/DFlash-DeepSeek-TraceMix/tree/v1.0.1) | Archived v1.0.1 release snapshot: trimmed text mix and stage-2 checkpoint. |
795
  | [`v1.0.0`](https://huggingface.co/datasets/biscet/DFlash-DeepSeek-TraceMix/tree/v1.0.0) | Archived v1.0.0 release snapshot with the full 50,000-task text mix. |
796
 
797
  The complete dataset is available from the default `main` branch and in the
798
- Dataset Viewer. Version branches preserve release snapshots; pass
799
- `revision="v1.0.1"` for this release or `revision="v1.0.0"` for the previous one.
800
- Agent records are train-only supplements, separate from the main held-out
801
- evaluation splits.
802
 
803
  ## The mix
804
 
805
  | Configuration | Focus | Train | Validation | Test |
806
  |:---|:---|---:|---:|---:|
807
- | `tasks` | Coding, text processing, knowledge and reasoning contexts | 37,690 | 2,152 | 2,145 |
808
- | `agent_steps` | One next-action step with tool history and observations | 294 | — | — |
809
- | `agent_trajectories` | Complete DeepSeek sessions with reasoning and tool use | 200 | — | — |
 
810
 
811
  The 294 steps are derived from the 200 trajectories. They are two views of the
812
  same agent sessions, not independent extra sessions.
813
 
814
- - **Main languages:** 20,685 Russian and 21,302 English tasks; no machine translations.
815
- - **Task mix:** 52.2% text/reasoning tasks (21,922) and 47.8% software development (20,065).
816
- - **Source mix:** 21,313 literary-text tasks, 19,354 anonymized project-code tasks,
817
  1,262 upstream prompt tasks and 58 self-created tasks.
818
- - **Splits:** 10,202 semantic families, with each family confined to one split.
819
  - **Agent traces:** 604 reasoning blocks and 611 tool calls across 200 teacher
820
  sessions, kept separate from the task contexts.
821
 
@@ -824,12 +827,12 @@ same agent sessions, not independent extra sessions.
824
  | Path | Contents |
825
  |:---|:---|
826
  | `data/train.jsonl`, `data/validation.jsonl`, `data/test.jsonl` | Task IDs, messages, source/provenance, rendered prompts and exact token lengths. |
827
- | `stage2/added_2026-11-13.jsonl` | The 272 tasks added in v1.0.1 as a standalone checkpoint. |
828
  | `agents/agent_trajectories.jsonl` | Sanitized complete DeepSeek trajectories. |
829
  | `agents/agent_steps.jsonl` | Sanitized next-action records compatible with the task schema. |
830
  | `capture/compact.jsonl` | Whittle capture inputs, at most 1,536 prompt tokens. |
831
  | `capture/extended.jsonl` | Longer Whittle capture inputs, 1,537–8,192 prompt tokens. |
832
- | `quality_report.json` | Main-corpus schema, split, duplicate, token-length and decontamination checks. |
833
  | `privacy_report.json` | Publication cleanup scope and automated privacy checks. |
834
  | `manifest.json` | Release counts, file sizes and SHA-256 checksums. |
835
  | `LICENSE`, `NOTICE` | Apache-2.0 release license and retained source attribution. |
@@ -845,7 +848,7 @@ tasks = load_dataset(
845
  agent_steps = load_dataset(
846
  "biscet/DFlash-DeepSeek-TraceMix", "agent_steps"
847
  )
848
- # v1.0.1 additions only
849
  stage2 = load_dataset(
850
  "biscet/DFlash-DeepSeek-TraceMix",
851
  data_files="stage2/added_2026-11-13.jsonl",
@@ -864,16 +867,15 @@ stage2 = load_dataset(
864
  logits or captured teacher token IDs. Generate the target model's responses
865
  and features through your capture pipeline before draft-model training.
866
  - **Teacher scope:** DeepSeek provenance applies to the agent configurations.
867
- The task-context configuration is not a corpus of DeepSeek completions.
868
- - **Lengths:** main-task prompts range from 52 to 12,967 tokens.
869
- 18 tasks exceed the 8,192-token capture budget: they stay in the task
870
- configs and the stage-2 checkpoint, but are not exported to `capture/`.
871
- Agent steps range from 1,128 to 8,176 tokens. Run compact and extended capture
872
- separately.
873
  - **Quality:** the saved main-corpus checks pass, with no detected exact or near
874
- duplicates across families and no cross-split 13-gram contamination detected in
875
- the added checkpoint records. Agent outputs are model-generated material,
876
- not verified reference answers or a held-out benchmark.
877
 
878
  </details>
879
 
@@ -890,10 +892,10 @@ project names were replaced with stable placeholders, repository file paths were
890
  replaced with hashed names, and commit revisions were dropped. Changed task
891
  prompts and token lengths were recalculated with the Whittle tokenizer.
892
 
893
- The publication audit checked 41,987 task rows plus the agent and capture
894
- copies, and found no remaining matches to the configured sensitive-data patterns.
895
- This is an automated pattern audit, not a guarantee that every possible
896
- identifying detail was removed.
897
 
898
  Raw logs, local configuration, checkpoints, reference solutions and rejected
899
  candidates are excluded. See the [privacy report](https://huggingface.co/datasets/biscet/DFlash-DeepSeek-TraceMix/blob/main/privacy_report.json).
 
729
 
730
  **DeepSeek teacher traces. Contexts for the next draft.**
731
 
732
+ `41,715 task contexts` &nbsp;·&nbsp; `272 stage-2 contexts` &nbsp;·&nbsp; `200 agent trajectories` &nbsp;·&nbsp; `294 next-action steps` &nbsp;·&nbsp; `RU / EN`
733
 
734
  [Explore the data](https://huggingface.co/datasets/biscet/DFlash-DeepSeek-TraceMix/tree/main) · [Manifest](https://huggingface.co/datasets/biscet/DFlash-DeepSeek-TraceMix/blob/main/manifest.json) · [Privacy report](https://huggingface.co/datasets/biscet/DFlash-DeepSeek-TraceMix/blob/main/privacy_report.json)
735
 
 
741
  > A bilingual mix of task prompts and DeepSeek-generated agent traces for
742
  > preparing draft-model training and studying tool use.
743
 
744
+ The main corpus contains **41,715 prompts and contexts**, not DeepSeek
745
  completions. The separate agent supplements contain actual reasoning, historical
746
  tool calls, real observations and final responses distilled from the DeepSeek
747
  teacher models listed below through Qwen Code.
 
750
 
751
  - **Text corpus trimmed.** Four complete public-domain literary sources were
752
  removed (8,285 tasks in total) to bring the text share down to about
753
+ one half of the main corpus.
754
+ - **New stage-2 contexts added.** 272 tasks built from two real codebases
755
  (`project_03`, `project_05_mobile`) were converted to the release schema,
756
+ anonymised and published as a standalone stage-2 checkpoint for the
757
+ **DeepSeek Flash** capture track.
758
+ - **Resulting main mix:** 52.6% text/reasoning and 47.4%
759
+ software development, with 20,546 Russian and 21,169 English tasks.
760
  - **Splits rebalanced** proportionally to 90 / 5 / 5.
761
 
762
  ## Stage-2 checkpoint (2026-11-13)
763
 
764
+ The 272 contexts added in this release are published as a standalone
765
  checkpoint, so they can be pulled out of the full dataset without filtering:
766
 
767
  [`stage2/added_2026-11-13.jsonl`](https://huggingface.co/datasets/biscet/DFlash-DeepSeek-TraceMix/blob/main/stage2/added_2026-11-13.jsonl)
768
 
769
+ - **272 records**, one JSON object per line, `split = "train"` in the file.
770
+ - These records are **not** duplicated inside `data/*.jsonl`; the main `tasks`
771
+ configuration keeps its original schema and split boundaries.
772
+ - They are prompt contexts prepared for the DeepSeek Flash capture stage, not
773
+ completed DeepSeek trajectories, so they are not added to the teacher
774
+ trajectory counts.
775
  - Every record carries `provenance.generator_version = "1.0.1+stage2-2026-11-13"`.
776
  - `source_sha256` is computed over the anonymised message/tool context, not over
777
  the original repository file.
 
 
 
 
778
 
779
  ## Teacher models
780
 
781
  We distilled the agent traces from **DeepSeek Flash v4.1 (1017)** and
782
  **DeepSeek Pro V4 (0813)**.
783
 
784
+ | Teacher model | Recorded API alias | Completed trajectories | v1.0.1 task contexts |
785
+ |:---|:---|---:|---:|
786
+ | DeepSeek Flash v4.1 (1017) | `deepseek-flash` | 161 | 272 |
787
+ | DeepSeek Pro V4 (0813) | `deepseek-v4-pro` | 39 | — |
788
 
789
+ The 272 new contexts are inputs for the DeepSeek Flash capture track;
790
+ they are not completed sessions, so the trajectory counts stay unchanged.
791
  Trace records retain the original API aliases in their model and provenance fields.
792
 
793
  ## Repository branches
794
 
795
  | Branch | Contents |
796
  |:---|:---|
797
+ | `main` | Current **v1.0.1** release: trimmed task contexts, stage-2 checkpoint, DeepSeek agent data, capture exports, manifests and reports. |
 
798
  | [`v1.0.0`](https://huggingface.co/datasets/biscet/DFlash-DeepSeek-TraceMix/tree/v1.0.0) | Archived v1.0.0 release snapshot with the full 50,000-task text mix. |
799
 
800
  The complete dataset is available from the default `main` branch and in the
801
+ Dataset Viewer. `v1.0.0` preserves the previous release snapshot; pass
802
+ `revision="v1.0.0"` for it. Agent records follow the same 90/5/5 split policy
803
+ by trajectory family and stay separate from the task splits.
 
804
 
805
  ## The mix
806
 
807
  | Configuration | Focus | Train | Validation | Test |
808
  |:---|:---|---:|---:|---:|
809
+ | `tasks` | Coding, text processing, knowledge and reasoning contexts | 37,500 | 2,110 | 2,105 |
810
+ | `agent_steps` | One next-action step with tool history and observations | 268 | 14 | 12 |
811
+ | `agent_trajectories` | Complete DeepSeek sessions with reasoning and tool use | 180 | 10 | 10 |
812
+ | `stage2` (file) | Stage-2 contexts for the DeepSeek Flash capture track | 272 | — | — |
813
 
814
  The 294 steps are derived from the 200 trajectories. They are two views of the
815
  same agent sessions, not independent extra sessions.
816
 
817
+ - **Main languages:** 20,546 Russian and 21,169 English tasks; no machine translations.
818
+ - **Task mix:** 52.6% text/reasoning tasks (21,922) and 47.4% software development (19,793).
819
+ - **Source mix:** 21,313 literary-text tasks, 19,082 anonymized project-code tasks,
820
  1,262 upstream prompt tasks and 58 self-created tasks.
821
+ - **Splits:** 10,076 semantic families, with each family confined to one split.
822
  - **Agent traces:** 604 reasoning blocks and 611 tool calls across 200 teacher
823
  sessions, kept separate from the task contexts.
824
 
 
827
  | Path | Contents |
828
  |:---|:---|
829
  | `data/train.jsonl`, `data/validation.jsonl`, `data/test.jsonl` | Task IDs, messages, source/provenance, rendered prompts and exact token lengths. |
830
+ | `stage2/added_2026-11-13.jsonl` | The 272 stage-2 contexts added in v1.0.1, split `train`. |
831
  | `agents/agent_trajectories.jsonl` | Sanitized complete DeepSeek trajectories. |
832
  | `agents/agent_steps.jsonl` | Sanitized next-action records compatible with the task schema. |
833
  | `capture/compact.jsonl` | Whittle capture inputs, at most 1,536 prompt tokens. |
834
  | `capture/extended.jsonl` | Longer Whittle capture inputs, 1,537–8,192 prompt tokens. |
835
+ | `quality_report.json` | Main-corpus schema, split, duplicate, token-length and privacy checks. |
836
  | `privacy_report.json` | Publication cleanup scope and automated privacy checks. |
837
  | `manifest.json` | Release counts, file sizes and SHA-256 checksums. |
838
  | `LICENSE`, `NOTICE` | Apache-2.0 release license and retained source attribution. |
 
848
  agent_steps = load_dataset(
849
  "biscet/DFlash-DeepSeek-TraceMix", "agent_steps"
850
  )
851
+ # v1.0.1 stage-2 checkpoint
852
  stage2 = load_dataset(
853
  "biscet/DFlash-DeepSeek-TraceMix",
854
  data_files="stage2/added_2026-11-13.jsonl",
 
867
  logits or captured teacher token IDs. Generate the target model's responses
868
  and features through your capture pipeline before draft-model training.
869
  - **Teacher scope:** DeepSeek provenance applies to the agent configurations.
870
+ The task-context configurations are not corpora of DeepSeek completions.
871
+ - **Lengths:** main-task prompts range from 52 to 5,030 tokens.
872
+ Stage-2 contexts are longer and are not part of `capture/`; the largest
873
+ stage-2 prompt is 12,967 tokens. Agent steps range from 1,128 to 8,176 tokens.
874
+ Run compact and extended capture separately.
 
875
  - **Quality:** the saved main-corpus checks pass, with no detected exact or near
876
+ duplicates across families. Stage-2 contexts are a separate supplement and
877
+ should be evaluated on their own split. Agent outputs are model-generated
878
+ material, not verified reference answers or a held-out benchmark.
879
 
880
  </details>
881
 
 
892
  replaced with hashed names, and commit revisions were dropped. Changed task
893
  prompts and token lengths were recalculated with the Whittle tokenizer.
894
 
895
+ The publication audit covered the task rows, the stage-2 checkpoint, the agent
896
+ records and the capture copies, and found no remaining matches to the configured
897
+ sensitive-data patterns. This is an automated pattern audit, not a guarantee that
898
+ every possible identifying detail was removed.
899
 
900
  Raw logs, local configuration, checkpoints, reference solutions and rejected
901
  candidates are excluded. See the [privacy report](https://huggingface.co/datasets/biscet/DFlash-DeepSeek-TraceMix/blob/main/privacy_report.json).
agents/agent_steps.jsonl CHANGED
The diff for this file is too large to render. See raw diff
 
agents/agent_trajectories.jsonl CHANGED
The diff for this file is too large to render. See raw diff
 
capture/compact.jsonl CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:1fe0339a34ce2417d6c079a9440123dcfe365eb28a78984ad7adf51a544b7995
3
- size 115703117
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:9124c78a080ec0856d4e75ab24c53b7a3a231cb92adb4b95e8efe94159b232ae
3
+ size 115627119
capture/extended.jsonl CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:9cfe6c49ab879edf55fbd1bc61cd82f7cddd05518ea974da80f47e9af5413a22
3
- size 14419864
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:56dc5dcd4486ed159c0ceff1994a6de4433731c4b4673d3ec0e8a915e1b780f5
3
+ size 10162310
data/test.jsonl CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:9eae2d2553bf587f476106aeef146a5d0f590c4bf84951e5aa57fd254244fb0f
3
- size 14134847
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:051024f1fc0ff64eb1fb79d2012d7f08dc7459f77737d04ea4b27730786c7182
3
+ size 12497371
data/train.jsonl CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:480223ea19007f6403f26c70fd4b2ed24bd349357dc3e08b817469cee5e9b990
3
- size 272386203
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:c7b886614647c1f67a8b0db45d2b34f6af875e599e4318118af9e7c1c97a1569
3
+ size 265345388
data/validation.jsonl CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:dd4bac9cbf4eccc4cb191768a4bc906efc202f3c2f8bb58780ea38cc36a1a966
3
- size 15074367
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:5b6e6eb5fa0b068f049ffc53510ec69b297e5d3b2e07257e2b410a38dbd8a78e
3
+ size 13349115
manifest.json CHANGED
@@ -1,5 +1,7 @@
1
  {
2
- "tasks": 41987,
 
 
3
  "trajectories": 200,
4
  "steps": 294,
5
  "files": {
@@ -16,52 +18,52 @@
16
  "bytes": 1689
17
  },
18
  "README.md": {
19
- "sha256": "b181f1c2868e572f54b5ba8ffcf101131668bbc50f188fd715dbe150fdce5c51",
20
- "bytes": 26241
21
  },
22
  "agents/agent_steps.jsonl": {
23
- "sha256": "8efb9d0b69c776ec5b93a04b115ecc0fc329170fd4a0fac1feea9f4ef2310be4",
24
- "bytes": 7026116
25
  },
26
  "agents/agent_trajectories.jsonl": {
27
- "sha256": "fcea0efedbcaf1ed785c61e8f7caf58ea175e76cd5ba84fc54e2ff582079041f",
28
- "bytes": 6539625
29
  },
30
  "capture/compact.jsonl": {
31
- "sha256": "1fe0339a34ce2417d6c079a9440123dcfe365eb28a78984ad7adf51a544b7995",
32
- "bytes": 115703117
33
  },
34
  "capture/extended.jsonl": {
35
- "sha256": "9cfe6c49ab879edf55fbd1bc61cd82f7cddd05518ea974da80f47e9af5413a22",
36
- "bytes": 14419864
37
  },
38
  "data/test.jsonl": {
39
- "sha256": "9eae2d2553bf587f476106aeef146a5d0f590c4bf84951e5aa57fd254244fb0f",
40
- "bytes": 14134847
41
  },
42
  "data/train.jsonl": {
43
- "sha256": "480223ea19007f6403f26c70fd4b2ed24bd349357dc3e08b817469cee5e9b990",
44
- "bytes": 272386203
45
  },
46
  "data/validation.jsonl": {
47
- "sha256": "dd4bac9cbf4eccc4cb191768a4bc906efc202f3c2f8bb58780ea38cc36a1a966",
48
- "bytes": 15074367
49
  },
50
  "privacy_report.json": {
51
- "sha256": "e9c1fa8843b83c8ba4a4584995d9a219f4cf4936f8e3e4daffe633e4c81248f3",
52
- "bytes": 1044
53
  },
54
  "quality_report.json": {
55
- "sha256": "cc5f69669e20896e9180c6d6faf90e4caa7c0fccb317c787d009a0e533f5e564",
56
- "bytes": 8717
57
  },
58
  "stage2/added_2026-11-13.jsonl": {
59
- "sha256": "a0d43442a57f943db01744aafad51e481b0259bc0094e4e7772d7d054078f476",
60
- "bytes": 10403543
61
  }
62
  },
63
  "version": "1.0.1",
64
  "repo_id": "biscet/DFlash-DeepSeek-TraceMix",
65
  "revision": "main",
66
- "snapshot_revision": "v1.0.1"
67
  }
 
1
  {
2
+ "tasks": 41715,
3
+ "stage2_tasks": 272,
4
+ "combined_tasks": 41987,
5
  "trajectories": 200,
6
  "steps": 294,
7
  "files": {
 
18
  "bytes": 1689
19
  },
20
  "README.md": {
21
+ "sha256": "9249da34dcf2e82b262bcff09e9e55f69ab3ef4df37170c14aa9d987c6936806",
22
+ "bytes": 26535
23
  },
24
  "agents/agent_steps.jsonl": {
25
+ "sha256": "e440c7fe9b33b4d80af209faab7020fc00847f9137c18cede38ad41bbc50a990",
26
+ "bytes": 7026174
27
  },
28
  "agents/agent_trajectories.jsonl": {
29
+ "sha256": "68c451c2fefe8b7042a1a9b3820f571cbee48a314c02c9efa498e308b311783a",
30
+ "bytes": 6539665
31
  },
32
  "capture/compact.jsonl": {
33
+ "sha256": "9124c78a080ec0856d4e75ab24c53b7a3a231cb92adb4b95e8efe94159b232ae",
34
+ "bytes": 115627119
35
  },
36
  "capture/extended.jsonl": {
37
+ "sha256": "56dc5dcd4486ed159c0ceff1994a6de4433731c4b4673d3ec0e8a915e1b780f5",
38
+ "bytes": 10162310
39
  },
40
  "data/test.jsonl": {
41
+ "sha256": "051024f1fc0ff64eb1fb79d2012d7f08dc7459f77737d04ea4b27730786c7182",
42
+ "bytes": 12497371
43
  },
44
  "data/train.jsonl": {
45
+ "sha256": "c7b886614647c1f67a8b0db45d2b34f6af875e599e4318118af9e7c1c97a1569",
46
+ "bytes": 265345388
47
  },
48
  "data/validation.jsonl": {
49
+ "sha256": "5b6e6eb5fa0b068f049ffc53510ec69b297e5d3b2e07257e2b410a38dbd8a78e",
50
+ "bytes": 13349115
51
  },
52
  "privacy_report.json": {
53
+ "sha256": "3847ac23d0f9814931e3e81302b47b1e4ad693e71182033f25ec04f3f2798026",
54
+ "bytes": 979
55
  },
56
  "quality_report.json": {
57
+ "sha256": "e0d0e10d135f02af3811964088934a4019865ca6c0a3d7c6f2c0bf6149f0fc72",
58
+ "bytes": 8783
59
  },
60
  "stage2/added_2026-11-13.jsonl": {
61
+ "sha256": "3cd22ac15af62e5edf348e9d2a3b5138f7bc61fb45f12361566ca5f2b70312ed",
62
+ "bytes": 10398717
63
  }
64
  },
65
  "version": "1.0.1",
66
  "repo_id": "biscet/DFlash-DeepSeek-TraceMix",
67
  "revision": "main",
68
+ "snapshot_revision": "main"
69
  }
privacy_report.json CHANGED
@@ -1,8 +1,8 @@
1
  {
2
- "checked_at": "2026-11-13T12:27:27+00:00",
3
  "release_version": "1.0.1",
4
  "all_checks_passed": true,
5
- "checked_jsonl_rows": 84722,
6
  "remaining_pattern_violations": 0,
7
  "violations_by_pattern": {},
8
  "allowed_placeholder_matches": {
@@ -16,12 +16,7 @@
16
  "rows": 272,
17
  "changed_rows": 272,
18
  "redactions": 1897
19
- },
20
- "data/train.jsonl": {
21
- "rows": 37690,
22
- "changed_rows": 272,
23
- "redactions": 1897
24
  }
25
  },
26
- "retained_audit": "Rows carried over from v1.0.0 were covered by the 2026-11-11 publication audit; no retained row was modified. New stage-2 rows were sanitised in this release."
27
  }
 
1
  {
2
+ "checked_at": "2026-11-13T12:36:56+00:00",
3
  "release_version": "1.0.1",
4
  "all_checks_passed": true,
5
+ "checked_jsonl_rows": 84196,
6
  "remaining_pattern_violations": 0,
7
  "violations_by_pattern": {},
8
  "allowed_placeholder_matches": {
 
16
  "rows": 272,
17
  "changed_rows": 272,
18
  "redactions": 1897
 
 
 
 
 
19
  }
20
  },
21
+ "retained_audit": "Rows carried over from v1.0.0 were covered by the 2026-11-11 publication audit; no retained row was modified, only four literary sources were removed. New stage-2 rows were sanitised in this release."
22
  }
quality_report.json CHANGED
@@ -3,72 +3,69 @@
3
  "target_model": "Whittle-Qwen-3.8-35B-A3B-Q5_K_M",
4
  "draft_architecture": "DFlash2",
5
  "seed": 641384055,
6
- "target_total_tasks": 41987,
7
- "completed_tasks": 41987,
8
  "completion_ratio": 1.0,
9
- "semantic_families_total": 10202,
10
  "splits": {
11
  "train": {
12
- "tasks": 37690,
13
- "semantic_families": 8490,
14
- "share_tasks": 0.897659,
15
- "share_families": 0.83219,
16
  "languages": {
17
- "en": 19152,
18
- "ru": 18538
19
  },
20
  "length_buckets": {
21
- "compact": 36245,
22
- "extended": 1433,
23
- "long": 12
24
  }
25
  },
26
  "validation": {
27
- "tasks": 2152,
28
- "semantic_families": 857,
29
- "share_tasks": 0.051254,
30
- "share_families": 0.084003,
31
  "languages": {
32
- "en": 1033,
33
- "ru": 1119
34
  },
35
  "length_buckets": {
36
- "compact": 2020,
37
- "extended": 128,
38
- "long": 4
39
  }
40
  },
41
  "test": {
42
- "tasks": 2145,
43
- "semantic_families": 855,
44
- "share_tasks": 0.051087,
45
- "share_families": 0.083807,
46
  "languages": {
47
- "ru": 1028,
48
- "en": 1117
49
  },
50
  "length_buckets": {
51
- "extended": 98,
52
- "compact": 2045,
53
- "long": 2
54
  }
55
  }
56
  },
57
  "language_distribution": {
58
- "en": 21302,
59
- "ru": 20685
60
  },
61
- "native_russian_count": 20685,
62
  "translated_russian_count": 0,
63
  "native_russian_ratio": 1.0,
64
  "source_kind_distribution": {
65
- "local_project": 19354,
66
  "public_domain_text": 21313,
67
  "open_dataset": 1262,
68
  "self_created": 58
69
  },
70
  "domain_distribution": {
71
- "frontend_ui": 18132,
72
  "technical_writing": 12989,
73
  "general_knowledge": 4484,
74
  "systems_programming": 19,
@@ -82,12 +79,12 @@
82
  "security_reliability": 9
83
  },
84
  "programming_language_distribution": {
85
- "TypeScript": 10860,
86
- "general": 22201,
87
  "C/C++": 23,
88
  "React": 2783,
89
- "JavaScript": 2083,
90
- "Vue": 1801,
91
  "Go": 1044,
92
  "HTML/CSS": 836,
93
  "SQL": 174,
@@ -97,91 +94,112 @@
97
  "Server/Infrastructure": 28
98
  },
99
  "task_type_distribution": {
100
- "api_and_database": 1887,
101
  "testing": 2592,
102
  "text_processing": 12989,
103
  "general_qa": 4486,
104
- "code_explanation": 3237,
105
- "refactoring": 3081,
106
- "bug_finding": 2768,
107
- "patch_preparation": 3130,
108
  "type_fixing": 2458,
109
  "logic_reasoning": 4251,
110
- "async_and_concurrency": 467,
111
  "function_implementation": 366,
112
- "agent_next_action": 33,
113
  "math_reasoning": 196,
114
  "multi_file_analysis": 14,
115
  "configuration": 26,
116
  "log_analysis": 6
117
  },
118
  "difficulty_distribution": {
119
- "hard": 5402,
120
- "medium": 19338,
121
- "easy": 17247
122
  },
123
  "context_size_distribution": {
124
- "long": 7729,
125
- "short": 16113,
126
- "medium": 18131,
127
  "multi_file": 14
128
  },
129
  "response_format_distribution": {
130
- "code_and_explanation": 6256,
131
  "test_suite": 1373,
132
  "concise_answer": 12464,
133
- "structured_analysis": 11424,
134
- "unified_diff_patch": 5544,
135
- "root_cause_and_fix": 2479,
136
  "step_by_step_solution": 2414,
137
- "tool_call": 33
138
  },
139
- "software_development_tasks": 20065,
140
- "software_development_share": 0.477886,
141
  "text_and_reasoning_tasks": 21922,
142
- "text_and_reasoning_share": 0.522114,
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
143
  "length_buckets": {
144
- "compact": 40310,
145
- "extended": 1659
146
  },
147
  "task_length_buckets": {
148
- "compact": 40310,
149
- "extended": 1659,
150
- "long": 18
151
  },
152
  "token_length_approximate": false,
153
  "length_statistics": {
154
  "all": {
155
- "count": 41987,
156
  "min_tokens": 52,
157
- "max_tokens": 12967,
158
- "mean_tokens": 613.69,
159
- "median_tokens": 454,
160
- "p95_tokens": 1460
161
  },
162
  "compact": {
163
- "count": 40310,
164
  "min_tokens": 52,
165
  "max_tokens": 1536,
166
- "mean_tokens": 547.44,
167
- "median_tokens": 447.0,
168
- "p95_tokens": 1232
169
  },
170
  "extended": {
171
- "count": 1659,
172
  "min_tokens": 1538,
173
- "max_tokens": 8150,
174
- "mean_tokens": 2127.76,
175
- "median_tokens": 1742,
176
- "p95_tokens": 4819
177
  },
178
  "long": {
179
- "count": 18,
180
- "min_tokens": 8197,
181
- "max_tokens": 12967,
182
- "mean_tokens": 9423.22,
183
- "median_tokens": 8963.0,
184
- "p95_tokens": 12918
185
  }
186
  },
187
  "duplicates_and_rejections": {
@@ -190,24 +208,20 @@
190
  "rejected_total": 0,
191
  "rejection_reasons": {},
192
  "same_family_variants_allowed": true,
193
- "unique_user_texts": 41987
194
  },
195
  "content_diversity_metrics": {
196
- "unique_source_groups": 10202,
197
- "unique_user_task_hashes": 41987,
198
- "unique_instruction_prefixes": 1086,
199
- "type_token_ratio": 0.013881,
200
  "max_family_size": 10,
201
- "mean_tasks_per_family": 4.116,
202
- "families_with_multiple_tasks": 8812
203
  },
204
  "privacy_and_decontamination": {
205
- "pre_tokenization_redactions": 12246,
206
  "hard_violations": {},
207
- "cross_split": {
208
- "heldout_exclusive_13grams": 217459,
209
- "contaminated_new_train_rows": 0
210
- },
211
  "secrets_or_pii_found": false
212
  },
213
  "release_changes": {
@@ -220,9 +234,7 @@
220
  "removed_text_tasks": 8285,
221
  "added_stage2_tasks": 272,
222
  "stage2_split_counts": {
223
- "train": 190,
224
- "validation": 42,
225
- "test": 40
226
  },
227
  "stage2_tag": "stage2-2026-11-13",
228
  "stage2_source_files": "stage2/added_2026-11-13.jsonl"
@@ -231,22 +243,22 @@
231
  {
232
  "check_name": "schema_valid",
233
  "passed": true,
234
- "details": "records=41987, unique_ids=True"
235
  },
236
  {
237
  "check_name": "split_family_disjoint",
238
  "passed": true,
239
- "details": "families=10202, multi_split_families=0"
240
  },
241
  {
242
  "check_name": "split_ratio_90_5_5",
243
  "passed": true,
244
- "details": "{\"test\": 2145, \"train\": 37690, \"validation\": 2152}"
245
  },
246
  {
247
  "check_name": "language_balance_45_55",
248
  "passed": true,
249
- "details": "{\"en\": 21302, \"ru\": 20685}"
250
  },
251
  {
252
  "check_name": "native_russian_share",
@@ -256,12 +268,12 @@
256
  {
257
  "check_name": "software_development_share",
258
  "passed": true,
259
- "details": "share=0.478"
260
  },
261
  {
262
  "check_name": "text_and_reasoning_share",
263
  "passed": true,
264
- "details": "share=0.522"
265
  },
266
  {
267
  "check_name": "no_assistant_completion",
@@ -301,7 +313,7 @@
301
  {
302
  "check_name": "length_buckets_consistent",
303
  "passed": true,
304
- "details": "{\"tasks\": {\"compact\": 40310, \"extended\": 1659, \"long\": 18}, \"capture\": {\"compact\": 40310, \"extended\": 1659}}"
305
  },
306
  {
307
  "check_name": "capture_covers_bounded_prompts",
@@ -311,7 +323,7 @@
311
  {
312
  "check_name": "prompt_length_bounds",
313
  "passed": true,
314
- "details": "min=52, max=12967"
315
  },
316
  {
317
  "check_name": "sources_present",
@@ -329,9 +341,9 @@
329
  "details": "converted=272, path_and_project_redactions=1897"
330
  },
331
  {
332
- "check_name": "cross_split_decontamination",
333
  "passed": true,
334
- "details": "{\"heldout_exclusive_13grams\": 217459, \"contaminated_new_train_rows\": 0, \"new_split_counts\": {\"train\": 190, \"validation\": 42, \"test\": 40}}"
335
  }
336
  ],
337
  "all_checks_passed": true
 
3
  "target_model": "Whittle-Qwen-3.8-35B-A3B-Q5_K_M",
4
  "draft_architecture": "DFlash2",
5
  "seed": 641384055,
6
+ "target_total_tasks": 41715,
7
+ "completed_tasks": 41715,
8
  "completion_ratio": 1.0,
9
+ "semantic_families_total": 10076,
10
  "splits": {
11
  "train": {
12
+ "tasks": 37500,
13
+ "semantic_families": 8398,
14
+ "share_tasks": 0.898957,
15
+ "share_families": 0.833466,
16
  "languages": {
17
+ "en": 19059,
18
+ "ru": 18441
19
  },
20
  "length_buckets": {
21
+ "compact": 36234,
22
+ "extended": 1266
 
23
  }
24
  },
25
  "validation": {
26
+ "tasks": 2110,
27
+ "semantic_families": 841,
28
+ "share_tasks": 0.050581,
29
+ "share_families": 0.083466,
30
  "languages": {
31
+ "en": 1013,
32
+ "ru": 1097
33
  },
34
  "length_buckets": {
35
+ "compact": 2016,
36
+ "extended": 94
 
37
  }
38
  },
39
  "test": {
40
+ "tasks": 2105,
41
+ "semantic_families": 837,
42
+ "share_tasks": 0.050461,
43
+ "share_families": 0.083069,
44
  "languages": {
45
+ "ru": 1008,
46
+ "en": 1097
47
  },
48
  "length_buckets": {
49
+ "extended": 60,
50
+ "compact": 2045
 
51
  }
52
  }
53
  },
54
  "language_distribution": {
55
+ "en": 21169,
56
+ "ru": 20546
57
  },
58
+ "native_russian_count": 20546,
59
  "translated_russian_count": 0,
60
  "native_russian_ratio": 1.0,
61
  "source_kind_distribution": {
62
+ "local_project": 19082,
63
  "public_domain_text": 21313,
64
  "open_dataset": 1262,
65
  "self_created": 58
66
  },
67
  "domain_distribution": {
68
+ "frontend_ui": 17860,
69
  "technical_writing": 12989,
70
  "general_knowledge": 4484,
71
  "systems_programming": 19,
 
79
  "security_reliability": 9
80
  },
81
  "programming_language_distribution": {
82
+ "TypeScript": 10642,
83
+ "general": 22157,
84
  "C/C++": 23,
85
  "React": 2783,
86
+ "JavaScript": 2077,
87
+ "Vue": 1797,
88
  "Go": 1044,
89
  "HTML/CSS": 836,
90
  "SQL": 174,
 
94
  "Server/Infrastructure": 28
95
  },
96
  "task_type_distribution": {
97
+ "api_and_database": 1859,
98
  "testing": 2592,
99
  "text_processing": 12989,
100
  "general_qa": 4486,
101
+ "code_explanation": 3085,
102
+ "refactoring": 3071,
103
+ "bug_finding": 2727,
104
+ "patch_preparation": 3108,
105
  "type_fixing": 2458,
106
  "logic_reasoning": 4251,
107
+ "async_and_concurrency": 460,
108
  "function_implementation": 366,
109
+ "agent_next_action": 21,
110
  "math_reasoning": 196,
111
  "multi_file_analysis": 14,
112
  "configuration": 26,
113
  "log_analysis": 6
114
  },
115
  "difficulty_distribution": {
116
+ "hard": 5181,
117
+ "medium": 19296,
118
+ "easy": 17238
119
  },
120
  "context_size_distribution": {
121
+ "long": 7538,
122
+ "short": 16111,
123
+ "medium": 18052,
124
  "multi_file": 14
125
  },
126
  "response_format_distribution": {
127
+ "code_and_explanation": 6221,
128
  "test_suite": 1373,
129
  "concise_answer": 12464,
130
+ "structured_analysis": 11272,
131
+ "unified_diff_patch": 5512,
132
+ "root_cause_and_fix": 2438,
133
  "step_by_step_solution": 2414,
134
+ "tool_call": 21
135
  },
136
+ "software_development_tasks": 19793,
137
+ "software_development_share": 0.474482,
138
  "text_and_reasoning_tasks": 21922,
139
+ "text_and_reasoning_share": 0.525518,
140
+ "stage2_tasks": 272,
141
+ "combined_tasks": 41987,
142
+ "stage2_language_distribution": {
143
+ "ru": 139,
144
+ "en": 133
145
+ },
146
+ "stage2_task_type_distribution": {
147
+ "patch_preparation": 22,
148
+ "api_and_database": 28,
149
+ "refactoring": 10,
150
+ "async_and_concurrency": 7,
151
+ "bug_finding": 41,
152
+ "code_explanation": 152,
153
+ "agent_next_action": 12
154
+ },
155
+ "agent_splits": {
156
+ "agent_trajectories": {
157
+ "validation": 10,
158
+ "train": 180,
159
+ "test": 10
160
+ },
161
+ "agent_steps": {
162
+ "validation": 14,
163
+ "train": 268,
164
+ "test": 12
165
+ }
166
+ },
167
  "length_buckets": {
168
+ "compact": 40295,
169
+ "extended": 1420
170
  },
171
  "task_length_buckets": {
172
+ "compact": 40295,
173
+ "extended": 1420
 
174
  },
175
  "token_length_approximate": false,
176
  "length_statistics": {
177
  "all": {
178
+ "count": 41715,
179
  "min_tokens": 52,
180
+ "max_tokens": 5030,
181
+ "mean_tokens": 589.07,
182
+ "median_tokens": 453,
183
+ "p95_tokens": 1423
184
  },
185
  "compact": {
186
+ "count": 40295,
187
  "min_tokens": 52,
188
  "max_tokens": 1536,
189
+ "mean_tokens": 547.19,
190
+ "median_tokens": 446,
191
+ "p95_tokens": 1231
192
  },
193
  "extended": {
194
+ "count": 1420,
195
  "min_tokens": 1538,
196
+ "max_tokens": 5030,
197
+ "mean_tokens": 1777.45,
198
+ "median_tokens": 1713.5,
199
+ "p95_tokens": 2183
200
  },
201
  "long": {
202
+ "count": 0
 
 
 
 
 
203
  }
204
  },
205
  "duplicates_and_rejections": {
 
208
  "rejected_total": 0,
209
  "rejection_reasons": {},
210
  "same_family_variants_allowed": true,
211
+ "unique_user_texts": 41715
212
  },
213
  "content_diversity_metrics": {
214
+ "unique_source_groups": 10076,
215
+ "unique_user_task_hashes": 41715,
216
+ "unique_instruction_prefixes": 1007,
217
+ "type_token_ratio": 0.014172,
218
  "max_family_size": 10,
219
+ "mean_tasks_per_family": 4.14,
220
+ "families_with_multiple_tasks": 8691
221
  },
222
  "privacy_and_decontamination": {
223
+ "pre_tokenization_redactions": 10349,
224
  "hard_violations": {},
 
 
 
 
225
  "secrets_or_pii_found": false
226
  },
227
  "release_changes": {
 
234
  "removed_text_tasks": 8285,
235
  "added_stage2_tasks": 272,
236
  "stage2_split_counts": {
237
+ "train": 272
 
 
238
  },
239
  "stage2_tag": "stage2-2026-11-13",
240
  "stage2_source_files": "stage2/added_2026-11-13.jsonl"
 
243
  {
244
  "check_name": "schema_valid",
245
  "passed": true,
246
+ "details": "records=41715, unique_ids=True"
247
  },
248
  {
249
  "check_name": "split_family_disjoint",
250
  "passed": true,
251
+ "details": "families=10076, multi_split_families=0"
252
  },
253
  {
254
  "check_name": "split_ratio_90_5_5",
255
  "passed": true,
256
+ "details": "{\"test\": 2105, \"train\": 37500, \"validation\": 2110}"
257
  },
258
  {
259
  "check_name": "language_balance_45_55",
260
  "passed": true,
261
+ "details": "{\"en\": 21169, \"ru\": 20546}"
262
  },
263
  {
264
  "check_name": "native_russian_share",
 
268
  {
269
  "check_name": "software_development_share",
270
  "passed": true,
271
+ "details": "share=0.474"
272
  },
273
  {
274
  "check_name": "text_and_reasoning_share",
275
  "passed": true,
276
+ "details": "share=0.526"
277
  },
278
  {
279
  "check_name": "no_assistant_completion",
 
313
  {
314
  "check_name": "length_buckets_consistent",
315
  "passed": true,
316
+ "details": "{\"tasks\": {\"compact\": 40295, \"extended\": 1420}, \"capture\": {\"compact\": 40295, \"extended\": 1420}}"
317
  },
318
  {
319
  "check_name": "capture_covers_bounded_prompts",
 
323
  {
324
  "check_name": "prompt_length_bounds",
325
  "passed": true,
326
+ "details": "min=52, max=5030"
327
  },
328
  {
329
  "check_name": "sources_present",
 
341
  "details": "converted=272, path_and_project_redactions=1897"
342
  },
343
  {
344
+ "check_name": "stage2_separate_from_tasks",
345
  "passed": true,
346
+ "details": "stage2_records=272, task_records=41715"
347
  }
348
  ],
349
  "all_checks_passed": true
stage2/added_2026-11-13.jsonl CHANGED
The diff for this file is too large to render. See raw diff