EuroswarmsInstitute/qwen3.8-max-glm5.2-kimi-k3-distillation-sua

qwen3.8-max-glm5.2-kimi-k3-distillation — System/User/Assistant format

Converted from r0b0tlab/qwen3.8-max-glm5.2-kimi-k3-distillation (canonical config, current shard set train-*-of-00006; the stale of-00005 shards in the source repo were excluded). Conversion date: 2026-08-20. License: inherited from the source — see LICENSE (controlled, noncommercial research scope).

Format

One JSON object per line, standard OpenAI-style chat format:

{"messages": [
  {"role": "system", "content": "..."},
  {"role": "user", "content": "..."},
  {"role": "assistant", "content": "<think>\n...\n</think>\n\n..."}
]}

• —Roles are only system / user / assistant; system (when present) is first, then strict user/assistant alternation, always ending on assistant.
• —Chain-of-thought: the source stores reasoning in a separate reasoning_content field. It is merged into the assistant content as a leading <think>...</think> block (46,600 assistant turns), matching the response format the source's system prompts mandate. Strip everything through </think>\n\n if you want final-answer-only targets.
• —Dropped source fields: trainable (true only on assistant turns — recover with assistant-only loss masking), sampling_weight, and all audit/metadata columns (id, domain, teacher_model, quality flags, provenance). Re-join via the source canonical config if needed; row order within each split is preserved.

Splits

Total 57,937 — every source trace is accounted for. 50,221 of the default traces are single-turn, 1,764 multi-turn.

Tool traces

5,952 traces (the source's Agent Tools domain) contain real tool-call trajectories — tool-role turns, tool_calls, tool_call_id and tool definitions — which cannot be expressed with only system/user/assistant roles. They are kept in the tool_traces config in OpenAI-style structured form: messages retain role, content, plus reasoning_content / tool_calls / tool_call_id / name when non-empty, and each record carries its tools definitions. Ignore this config for plain chat SFT.

Usage

from datasets import load_dataset

ds = load_dataset("EuroswarmsInstitute/qwen3.8-max-glm5.2-kimi-k3-distillation-sua")
# ds["train"][0]["messages"] -> [{"role": "system", ...}, {"role": "user", ...}, {"role": "assistant", ...}]

The JSONL files load directly in Unsloth Studio (auto-detected as the OpenAI-style messages format) and TRL's SFTTrainer conversational format.

Caveats

• —Validation/test are development holdouts, benchmark-derived per the source — do not report them as capability evaluations.
• —The <think>-formatted targets suit reasoning/thinking models natively; on non-reasoning models they will teach think-block emission.
• —Uniform sampling only: the source's sft_balanced weights are not carried into this view.