satgeze/Qwen3.6-27B-selfdistill
Qwen3.6-27B self-distillation pairs (DSpark drafter training data) The exact training data behind satgeze/Qwen3.6-27B-DSpark (2.5-2.7x measured decode speedup in llama.cpp): public prompts answered by the target model itself, so the drafter trains on precisely the distribution it drafts for at inference. The responses were generated by Qwen3.6-27B and exist nowhere upstream. Provenance, exactly part source license prompts mlabonne/open-perfectblend (12K… See the full description on the dataset page: https://huggingface.co/datasets/satgeze/Qwen3.6-27B-selfdistill.
Qwen3.6-27B self-distillation pairs (DSpark drafter training data)
The exact training data behind satgeze/Qwen3.6-27B-DSpark (2.5-2.7x measured decode speedup in llama.cpp): public prompts answered by the target model itself, so the drafter trains on precisely the distribution it drafts for at inference. The responses were generated by Qwen3.6-27B and exist nowhere upstream.
Provenance, exactly
Generation: llama-server (bf16, parallel slots) via satindergrewal/DeepSpec:qwen35-lane generate_train_data.py; temperature 0.7, thinking disabled, full-length responses (avg ~3,100 chars). Format: DeepSpec-standard JSONL, conversations field.
File
Sibling dataset: satgeze/Qwen3.5-0.8B-selfdistill (same recipe on the 0.8B testbed, including the truncated-vs-full-length supervision-volume comparison pair).
Reproduce
Full pipeline (regenerate, capture from GGUF, train, convert, serve): gguf-capture-training.md