SauravMahalik/varek-sft-mix-a-general-coding

Varek SFT Mix A: Generalized Coding Corpus

Part of the Varek research program on edge-deployable sub-1B coding models.

Dataset description

This repository contains 89,875 curated conversations produced for supervised fine-tuning of the Varek model family. The corpus merges multi-teacher distillation data and real-world agent traces spanning software engineering, mathematics, reasoning, tool use, and security-oriented coding, then applies cross-source deduplication, quality filtering, and strict evaluation holdout separation.

Composition

Source shard	Upstream dataset	License	Conversations
qwen3.8-max-50k	saidutta69/qwen3.8-max-distillation-50k-clean	Other (upstream terms)	44,179
qgk-sft_code	saidutta69/qwen-glm-kimi-distillation-clean	Other (upstream terms)	13,427
qgk-sft_reasoning	saidutta69/qwen-glm-kimi-distillation-clean	Other (upstream terms)	9,950
multi-distilled	saidutta69/GPT-5.5-Gemini-3.1-Pro-Grok-4-Claude-Fable-Mythos-5-Qwen-3.7-Max-Distillation-Cleaned	Custom (upstream terms)	8,242
qgk-sft_math	saidutta69/qwen-glm-kimi-distillation-clean	Other (upstream terms)	4,891
multi-cybersecurity	saidutta69/GPT-5.5-Gemini-3.1-Pro-Grok-4-Claude-Fable-Mythos-5-Qwen-3.7-Max-Distillation-Cleaned	Custom (upstream terms)	4,849
multi-coding	saidutta69/GPT-5.5-Gemini-3.1-Pro-Grok-4-Claude-Fable-Mythos-5-Qwen-3.7-Max-Distillation-Cleaned	Custom (upstream terms)	2,229
kimik3	saidutta69/kimi-k3-distillation-clean	Other (upstream terms)	2,033
claude-mythos	saidutta69/claude-mythos-distilled-25k-clean	Apache-2.0	75

Total: 89,875 conversations across domains including 50k, coding, cybersecurity, distilled, kimik3, mythos, sft_code, sft_math, sft_reasoning.

Processing pipeline

1. Decoding of serialized messages fields from upstream parquet storage.
2. Normalization of OpenAI-format tool calls (arguments JSON strings converted to objects) for compatibility with standard chat templates.
3. Quality gating: minimum quality scores enforced per source where quality metadata exists; zero-weight records removed.
4. Cross-source deduplication via exact-conversation SHA-256 and assistant-output-only SHA-256 hashing; first occurrence retained in priority order of curation quality.
5. Evaluation integrity: test splits of every upstream family are globally blocklisted so that no held-out evaluation item appears in training data.
6. Exclusion policy: RaceBench is excluded entirely and was never loaded during preparation.

Data schema

Field	Type	Description
messages	string (JSON)	Conversation as an array of {role, content, tool_calls?} objects
tools	string (JSON) or null	Tool schemas available during the conversation
source_dataset	string	Provenance shard identifier
domain	string	Coarse task domain label
teacher_model	string or null	Teacher model that produced the response
quality	float or null	Source-specific quality score
n_messages	int	Message count

The dataset ships as a single training split in parquet format.

Intended use

Supervised fine-tuning of small decoder-only language models for generalized coding, tool-assisted development workflows, and structured reasoning. Not intended for redistribution of individual trace content or for re-identification of contributors in anonymized trace sources.

Limitations

• Distilled outputs inherit teacher-model biases and may contain factual errors.
• Domain balance reflects curation priorities rather than natural code distribution.
• Python-centric coverage; multilingual code representation is limited.
• The multi-* shards derive from a custom-licensed upstream; commercial use should be reviewed against those terms.

Licensing and attribution

This derivative aggregates multiple upstream datasets under differing licenses (MIT, Apache-2.0, and custom upstream terms). The composite release is therefore marked license: other. Per-source licensing is tabulated above. Build pipeline: deterministic, seeded (42), reproducible from the upstream repositories.