biscet/DFlash-DeepSeek-TraceMix

DFlash DeepSeek TraceMix

DeepSeek teacher traces. Contexts for the next draft.

41,715 task contexts  ·  272 stage-2 contexts  ·  200 agent trajectories  ·  294 next-action steps  ·  RU / EN

Explore the data · Manifest · Privacy report

---

Built for DFlash distillation experiments. A bilingual mix of task prompts and DeepSeek-generated agent traces for preparing draft-model training and studying tool use.

The main corpus contains 41,715 prompts and contexts, not DeepSeek completions. The separate agent supplements contain actual reasoning, historical tool calls, real observations and final responses distilled from the DeepSeek teacher models listed below through Qwen Code.

What changed in v1.0.1

• Text corpus trimmed. Four complete public-domain literary sources were removed (8,285 tasks in total) to bring the text share down to about one half of the main corpus.
• New stage-2 contexts added. 272 tasks built from two real codebases (project_03, project_05_mobile) were converted to the release schema, anonymised and published as a standalone stage-2 checkpoint for the DeepSeek Flash capture track.
• Resulting main mix: 52.6% text/reasoning and 47.4% software development, with 20,546 Russian and 21,169 English tasks.
• Splits rebalanced proportionally to 90 / 5 / 5.

Stage-2 checkpoint (2026-11-13)

The 272 contexts added in this release are published as a standalone checkpoint, so they can be pulled out of the full dataset without filtering:

stage2/added_2026-11-13.jsonl

• 272 records, one JSON object per line, split = "train" in the file.
• These records are not duplicated inside data/*.jsonl; the main tasks configuration keeps its original schema and split boundaries.
• They are prompt contexts prepared for the DeepSeek Flash capture stage, not completed DeepSeek trajectories, so they are not added to the teacher trajectory counts.
• Every record carries provenance.generator_version = "1.0.1+stage2-2026-11-13".
• source_sha256 is computed over the anonymised message/tool context, not over the original repository file.

Teacher models

We distilled the agent traces from DeepSeek Flash v4.1 (1017) and DeepSeek Pro V4 (0813).

Teacher model	Recorded API alias	Completed trajectories	v1.0.1 task contexts
DeepSeek Flash v4.1 (1017)	deepseek-flash	161	272
DeepSeek Pro V4 (0813)	deepseek-v4-pro	39	—

The 272 new contexts are inputs for the DeepSeek Flash capture track; they are not completed sessions, so the trajectory counts stay unchanged. Trace records retain the original API aliases in their model and provenance fields.

Repository branches

Branch	Contents
main	Current v1.0.1 release: trimmed task contexts, stage-2 checkpoint, DeepSeek agent data, capture exports, manifests and reports.
v1.0.0	Archived v1.0.0 release snapshot with the full 50,000-task text mix.

The complete dataset is available from the default main branch and in the Dataset Viewer. v1.0.0 preserves the previous release snapshot; pass revision="v1.0.0" for it. Agent records follow the same 90/5/5 split policy by trajectory family and stay separate from the task splits.

The mix

Configuration	Focus	Train	Validation	Test
tasks	Coding, text processing, knowledge and reasoning contexts	37,500	2,110	2,105
agent_steps	One next-action step with tool history and observations	268	14	12
agent_trajectories	Complete DeepSeek sessions with reasoning and tool use	180	10	10
stage2 (file)	Stage-2 contexts for the DeepSeek Flash capture track	272	—	—

The 294 steps are derived from the 200 trajectories. They are two views of the same agent sessions, not independent extra sessions.

• Main languages: 20,546 Russian and 21,169 English tasks; no machine translations.
• Task mix: 52.6% text/reasoning tasks (21,922) and 47.4% software development (19,793).
• Source mix: 21,313 literary-text tasks, 19,082 anonymized project-code tasks, 1,262 upstream prompt tasks and 58 self-created tasks.
• Splits: 10,076 semantic families, with each family confined to one split.
• Agent traces: 604 reasoning blocks and 611 tool calls across 200 teacher sessions, kept separate from the task contexts.

Files

Path	Contents
data/train.jsonl, data/validation.jsonl, data/test.jsonl	Task IDs, messages, source/provenance, rendered prompts and exact token lengths.
stage2/added_2026-11-13.jsonl	The 272 stage-2 contexts added in v1.0.1, split train.
agents/agent_trajectories/{train,validation,test}.jsonl	Sanitized complete DeepSeek trajectories.
agents/agent_steps/{train,validation,test}.jsonl	Sanitized next-action records compatible with the task schema.
capture/compact.jsonl	Whittle capture inputs, at most 1,536 prompt tokens.
capture/extended.jsonl	Longer Whittle capture inputs, 1,537–8,192 prompt tokens.
quality_report.json	Main-corpus schema, split, duplicate, token-length and privacy checks.
privacy_report.json	Publication cleanup scope and automated privacy checks.
manifest.json	Release counts, file sizes and SHA-256 checksums.
LICENSE, NOTICE	Apache-2.0 release license and retained source attribution.

Load the release

from datasets import load_dataset

tasks = load_dataset(
    "biscet/DFlash-DeepSeek-TraceMix", "tasks"
)
agent_steps = load_dataset(
    "biscet/DFlash-DeepSeek-TraceMix", "agent_steps"
)
# v1.0.1 stage-2 checkpoint
stage2 = load_dataset(
    "biscet/DFlash-DeepSeek-TraceMix",
    data_files="stage2/added_2026-11-13.jsonl",
    split="train",
)

Before your training run

• Target format: task prompts use the chat template for Whittle-Qwen-3.8-35B-A3B-Q5_K_M, with enable_thinking=false. They contain no assistant answers for ordinary tasks. Agent records can contain historical assistant tool calls and actual tool observations.
• DFlash capture: this release does not contain target-model hidden features, logits or captured teacher token IDs. Generate the target model's responses and features through your capture pipeline before draft-model training.
• Teacher scope: DeepSeek provenance applies to the agent configurations. The task-context configurations are not corpora of DeepSeek completions.
• Lengths: main-task prompts range from 52 to 5,030 tokens. Stage-2 contexts are longer and are not part of capture/; the largest stage-2 prompt is 12,967 tokens. Agent steps range from 1,128 to 8,176 tokens. Run compact and extended capture separately.
• Quality: the saved main-corpus checks pass, with no detected exact or near duplicates across families. Stage-2 contexts are a separate supplement and should be evaluated on their own split. Agent outputs are model-generated material, not verified reference answers or a held-out benchmark.

Privacy and publication processing

Personal paths, configured private identifiers, internal project names, email addresses, private IPs, phone numbers, authorization values and known token/key patterns are redacted across JSON keys and values, including reasoning and tool arguments. Session identifiers are removed from normalized agent records.

The v1.0.1 stage-2 records were additionally anonymised before publication: project names were replaced with stable placeholders, repository file paths were replaced with hashed names, and commit revisions were dropped. Changed task prompts and token lengths were recalculated with the Whittle tokenizer.

The publication audit covered the task rows, the stage-2 checkpoint, the agent records and the capture copies, and found no remaining matches to the configured sensitive-data patterns. This is an automated pattern audit, not a guarantee that every possible identifying detail was removed.

Raw logs, local configuration, checkpoints, reference solutions and rejected candidates are excluded. See the privacy report.

Attribution & terms

Release license: Apache-2.0. Original contributions and release packaging are provided under this license. Third-party source terms remain applicable; the license does not replace source notices recorded in each task's source and provenance fields.

Sources include anonymized project-code excerpts, public-domain literary passages from Project Gutenberg and Wikisource, self-created tasks and prompts from biscet/DFlash-TraceMix-Lite v1.0.0. Tulu-related sources retain ODC-BY-1.0 and subset-specific terms. Literary source records retain book titles, authors, URLs and license notices. Agent records retain the DeepSeek model and Qwen Code generator provenance.

See NOTICE for attribution and source terms.