Hemmingway-1 Heretic
Altworld/Hemmingway-1 (a Qwen3.8-27B fine-tune for everyday writing) with its refusal behaviour removed by Heretic using full-weight Arbitrary-Rank Ablation: 0/100 refusals (original: 98/100) at KL divergence 0.083. Reasoning, tool calling, vision and fine-tunability are all checked below. All credit for the model goes to Altworld.
Reduced safety guardrails by design. You are responsible for what you do with it.
| Repository | Format |
|---|---|
| this repo | bf16 safetensors, stock Qwen3.8 vision-language layout (Qwen3_5ForConditionalGeneration) |
| Hemmingway-1-Heretic-Text | bf16 safetensors, text-only layout (Qwen3_5ForCausalLM), the easiest base for text/tool-call fine-tuning |
| Hemmingway-1-Heretic-GGUF | GGUF BF16 / Q8_0 / Q6_K / Q4_K_M + vision mmproj, each tested (KLD, top-token, refusals, tool calls) |
| Hemmingway-1-Heretic-FP8 | FP8 W8A8 compressed-tensors for vLLM, 35 GB |
| Hemmingway-1-Heretic-NVFP4 | NVFP4 compressed-tensors for vLLM on Blackwell, 27 GB |
Results
Refusals on 100 held-out harmful prompts (mlabonne/harmful_behaviors, strict refusal-marker list). KL divergence of
first-token distributions on 100 harmless prompts vs the original, measured by Heretic in a separate run on the saved
model:
| Refusals | KL divergence | |
|---|---|---|
| Hemmingway-1 (original) | 98/100 | 0 |
| Hemmingway-1 Heretic (this) | 0/100 | 0.083 |
| JohnDi/Hemmingway-1-Abliterated-Extreme, for reference | 2/100 | 0.187 |
Reasoning (thinking mode, greedy, the same problems for both models):
| GSM8K (40) | MATH-500 levels 4-5 (40) | |
|---|---|---|
| Hemmingway-1 | 40/40 | 32/40 |
| Heretic | 40/40 | 33/40 |
On MATH both solve the same 31 problems. The original alone solves 1, Heretic alone 2: no measurable loss.
Fine-tuning (40-step LoRA, r=16, on persona + tool-call chats): loss 0.96 → 0.004, the adapter reloads, and the
tuned model answers an unseen question with a well-formed <tool_call>. It merges back into full weights.
How it was made
- Hemmingway-1 ships as text-only safetensors. It was first grafted into the stock Qwen3.8-27B layout (vision tower restored). That conversion was verified byte-identical and logit-identical: darrellbest/Hemmingway-1-VL.
- Heretic (full-weight ARA:
use_ara = true,use_ara_lora = false). The search was seeded with parameter sets that had already worked on the same base model: our own Qwen3.8-27B release and trohrbaugh/Qwen3.8-27B-heretic-ara. The chosen trial is trohrbaugh's set: layers 26-56, preserve 0.9432, steer 0.0009, overcorrect 0.5038, neighbours 10. - Only 60 tensors changed, the output projections (
mlp.down_proj,self_attn.o_proj,linear_attn.out_proj) of layers 26-55. Embeddings, the vision tower and all other weights are byte-identical to Hemmingway-1. - The multi-token-prediction head (
mtp.*), whichsave_pretraineddrops, was restored unchanged.
Use
from transformers import AutoModelForImageTextToText, AutoTokenizer
tok = AutoTokenizer.from_pretrained("darrellbest/Hemmingway-1-Heretic")
model = AutoModelForImageTextToText.from_pretrained("darrellbest/Hemmingway-1-Heretic", dtype="auto", device_map="auto")
vllm serve darrellbest/Hemmingway-1-Heretic
Licence
Same as Hemmingway-1: CC BY-NC 4.0. Non-commercial use, with credit to Hemmingway-1 / Altworld. Commercial use needs an agreement with Altworld (luka@hemmingway.io).