Most forks try to change assistant style with adapters or prompt wrappers; this release provides a full-weight fine-tune that deliberately shifts Qwen3.6-27B toward stepwise, trace-like assistant behavior. That shift is intended to make outputs more guided and explanatory for code and technical-reasoning workflows while remaining suitable for local conversion and inference.
Key Capabilities
- Instruction + trace-style fine-tuning: trained on a cleaned Fable 5-style dataset to encourage step-by-step, explainable responses rather than terse answers. This means prompts that request reasoning traces or debugging steps tend to produce more structured outputs.
- Full HF checkpoint (no PEFT/LoRA): the repository contains the full model weights rather than an adapter, so behavior changes are baked into the checkpoint and will persist after conversion to local formats.
- Local-first conversion notes: explicitly prepared for downstream conversion to GGUF for runtimes like llama.cpp or LM Studio; the checkpoint has no MTP layers (MTP = 0) and includes guidance for validation after conversion.
- Practical focus: tuned toward code generation/editing, technical debugging assistance, and reasoning-heavy prompts rather than general-purpose chat persona or multimodal tasks.
Who It's For and Trade-offs
Great fit if you want a self-hostable, instruction-tuned Qwen3.6-27B variant that favors explicit reasoning traces and structured technical answers—useful for offline coding assistants, local agent experiments, and when you plan to convert to GGUF. Look elsewhere if you need a lightweight adapter (LoRA), an MTP-trained variant, or a model already validated for high-stakes production; this checkpoint may inherit base-model limitations, can produce confident but incorrect answers, and requires user validation after conversion.
Where It Fits
Compared with adapter-based tweaks, Qwable 27B trades smaller distribution size for a fully integrated behavioral shift: it's heavier to distribute but simpler to run as a standalone checkpoint. Compared with the untouched base, expect differences in refusal behavior, style, and safety posture because behavior is modified at weight-level fine-tuning.
How It Was Fine-tuned
The run used a normalized chat-format JSONL training set (data/processed/train.jsonl) in a DSv4-Tune workflow; typical guidance is to verify tokenizer/base revision and validate generation outputs after any export or conversion. The model card explicitly recommends test prompts (explain the model in 3 short paragraphs, write a JSONL validator script, debug a broken Docker Compose file) to confirm correct conversion and generation behavior.
