ThinkingCap-Qwen3.6-27B-Heretic-FABLEANT — Heretic Abliterated + FABLE.5 Traces LoRA

Uncensored reasoning model. Built from bottlecapai/ThinkingCap-Qwen3.6-27B, abliterated with Heretic to remove refusal behavior, then merged with a LoRA trained on the Complete-FABLE.5-traces-2M reasoning-trace dataset. Quantized to Q4_K_M GGUF.

Quick start

ollama run n0404n0404/thinkingcap-qwen3.6-27b-heretic-6f87a8-fableant:Q4_K_M

This is a thinking model — it reasons in a separate thinking channel before answering. When using the API, make sure the generation budget is large enough (e.g. num_predict ≥ 2048), otherwise the reasoning phase can consume the entire budget and the final answer gets cut off.

Build pipeline

1. Base: bottlecapai/ThinkingCap-Qwen3.6-27B (Qwen3.6-27B reasoning model)
2. Abliteration: Heretic directional ablation, 200 Optuna trials (TPE-optimized per-layer weights, orthogonalized direction, full row normalization), best trial merged into full weights
3. LoRA merge: a LoRA fine-tuned on Complete-FABLE.5-traces-2M (2M reasoning traces) merged into the abliterated model to restore/strengthen reasoning quality
4. Quantization: converted to GGUF and quantized to Q4_K_M with llama.cpp

Benchmarks

gsm8k
  sample_len            1319.000
  exact_match           97.12 %   (flexible-extract)
  exact_match           97.19 %   (strict-match)

Full GSM8K test set (all 1319 questions, 5-shot), measured on this exact Q4_K_M GGUF served by Ollama — not the fp16 safetensors.

Evaluation tooling: EleutherAI LM Evaluation Harness (lm-eval), using the local-chat-completions backend against Ollama’s OpenAI-compatible API, with the chat template applied and thinking enabled. Default completion-style stop strings were disabled and the generation budget raised so the reasoning phase is never truncated — without this, thinking models score near 0 on generative benchmarks regardless of actual capability.

Abliteration metrics (from the Heretic run, evaluated against the original base model):

The very low KL divergence means behavior on normal prompts is nearly unchanged — and the full-testset GSM8K score confirms reasoning capability survived abliteration, LoRA merging, and 4-bit quantization intact.

Intended use & disclaimer

Abliteration removes the model’s tendency to refuse requests. It does not add new knowledge or capabilities — it only stops the model from declining. This model is intended for research, red-teaming, creative writing, and other lawful uses where refusals get in the way.

You are responsible for how you use it. Outputs are not reviewed; do not use this model for illegal purposes or to cause harm.

License

Follows the license of the base model (bottlecapai/ThinkingCap-Qwen3.6-27B, Qwen3.6 series).

中文說明

本模型以 bottlecapai/ThinkingCap-Qwen3.6-27B 為基底,經 Heretic 消融(abliteration)移除拒答行為(200 次 Optuna 試驗取最佳解:關鍵字拒答 ⁹⁷⁄₁₀₀ → ¹⁹⁄₁₀₀,KL 散度僅 0.021),再合併以 Complete-FABLE.5-traces-2M 推理軌跡資料集訓練的 LoRA,最後量化為 Q4_K_M GGUF。

GSM8K 完整測試集實測(本 Q4 量化版、經 Ollama 服務):sample_len 1319、exact_match 97.12%(flexible-extract)/ 97.19%(strict-match),5-shot。評測工具為 EleutherAI LM Evaluation Harness(lm-eval),透過 Ollama 的 OpenAI 相容 API 以 local-chat-completions 後端執行,套用聊天模板並開啟思考模式。

這是思考型模型,API 使用時請將 num_predict 設 2048 以上,避免思考過程耗盡生成額度導致答案被截斷。請合法、負責任地使用。