The dataset viewer is not available because its heuristics could not detect any supported data files. You can try uploading some data files, or configuring the data files location manually.
Qwen3.6 Table2 80% + SynthDoc self-reflection 20% training bundle
| field | value |
|---|---|
experiment |
One-epoch Qwen3.6-27B assistant-only LoRA SFT, mixing filtered Table-2 instruction data with first-person SynthDoc self-reflection at 80/20 by loss-bearing tokens. |
date_generated |
2026-08-04 |
constitution |
constitutions/claude_distilled_12_principles_mid/constitution.md; both upstream corpora connect to this target. |
source_repo |
Matthew-Bozoukov/teaching_claude_why_replication commit 2de376e1e0f4a6b783b1e5ae98091b54aba1b7e7, branch codex/sft-self-reflect-20-table2-80. |
models |
Base Qwen/Qwen3.6-27B revision 6a9e13bd6fc8f0983b9b99948120bc37f49c13e9. |
generation_config |
No generation was performed; deterministic source shuffle seed 0 and greedy token-budget fill. |
schema |
mixture.jsonl: text is preserved-think Qwen3.6 ChatML; source is table2_filtered or synthdoc_self_reflection. |
provenance |
uv run scripts/data/build_mixture.py --config configs/data/mixture_qwen36_table2_80_synthdoc_self_reflect_20.yaml; exact code is code.tar.gz. |
Exact mixture
- 5,632 examples; 2,662,269 rendered tokens.
- 1,492,977 supervised tokens: 1,194,418 Table2 (80.00%) and 298,559 SynthDoc (20.00%).
- Source revisions and exact filenames are pinned in
data_config.yaml. - Mixture SHA-256:
7bb8a9b3bd2eaae6886592d9962e4733bec915f0f9374dd99e9bc1888a9569bc.
The loss mask excludes system/user text, assistant headers, and Qwen3.6's forced <think>
prefill. Real reasoning traces, their generated close, visible answers, and <|im_end|> are
supervised. On no-reasoning Table2 turns the full forced empty marker is masked. The independent
pre-train mask gate passed: 64 rows decoded exactly, with 260 real, 5,613 empty, and 0 absent think
blocks across the whole mixture.
Training recipe
Full-BF16 LoRA (not 4-bit QLoRA): r=32, alpha=64, dropout=0.05, language-model attention/MLP
projection targets, one epoch, batch 1 x accumulation 16, LR 1e-4 cosine, 3% warmup, SDPA,
gradient checkpointing, no packing, max sequence length 4096. See train_config.yaml.
- Downloads last month
- -