Downloads · 30 days
159
22% of all-time downloads
brainnxdomain/Hemmingway-1-Heretic-MTP-GGUF-V2
Hemmingway-1-Heretic-MTP-GGUF-V2 is a machine learning model from brainnxdomain. Use it for the machine learning task on the model card, and read the license before you ship it in a product. The card lists the license as cc-by-nc-4.0.
Hemmingway-1-Heretic-MTP-V3-Final-GGUF stays decensored with thinking on (2 vs 60–87 per 100). Keep V2 only for thinking-off serving. This repo remains for reference only.
Downloads · 30 days
159
22% of all-time downloads
All-time downloads
725
Public
Repo size
142 GB
Likes
1
Public
Click a slice to open those files.
.gguf142 GB · 100%
From the Hugging Face model README
Hemmingway-1-Heretic-MTP-V3-Final-GGUF stays decensored with thinking on (2 vs 60–87 per 100). Keep V2 only for thinking-off serving. This repo remains for reference only.
V2 is the second release of our decensored Hemmingway-1. V1 (Hemmingway-1-Heretic-MTP-GGUF, trial 106) used the ARA branch of Heretic. V2 (trial 208) uses the traditional Heretic method with a warm-started search. Both releases keep Hemmingway-1's writing ability as their goal. V2 does the job better on every measure we ran.
Base model: Altworld/Hemmingway-1 — 27B, Qwen3.8 lineage, 64 decoder layers plus one MTP block, 262,144-token context.
V1 removed refusals but paid for it in instruction following. V2 cuts that cost by roughly seven times and removes more refusals.
| Measure | Original | V1 (trial 106) | V2 (trial 208) |
|---|---|---|---|
| Refusals, non-thinking (Heretic test) | 97/100 | 7/100 | 4/100 |
| KL divergence from original | 0 | 0.0576 | 0.0112 |
| Constraint adherence, thinking off | 91.7% | 33.3% | 83.3% |
| Response length, thinking on | baseline | −74.5% | −47.1% |
| Constraint adherence, thinking on | 58.3% | 25.0% | 50.0% |
| EQ-Bench (±1.4) | 83.17 | 83.50 | 83.22 |
| HellaSwag acc / acc_norm (100 subset) | 59% / 76% | 59% / 76% | 59% / 76% |
EQ-Bench and HellaSwag tie with the original in all three cases. The difference is the middle rows: V2 stays close to the original where V1 drifted hard. In our 12-prompt creative suite, V1 ignored stated constraints in two of three responses; V2 ignores them in one of six.
If you use V1 today, switch. Nothing we measured favors it.
You are a helpful assistant.All quants include the MTP block:
| File | Size |
|---|---|
| BF16 | 54.7 GB |
| Q8_0 | 29.0 GB |
| Q6_K | 22.4 GB |
| Q5_K_M | 19.5 GB |
| Q4_K_M | 16.8 GB |
Use a recent build with Qwen3.5 and MTP support. The chat template is embedded; add --jinja.
llama-server -m Hemmingway-1-Heretic-V2-Q5_K_M.gguf \
--jinja -fa \
-c 32768 --temp 0.7 --min-p 0.1 --top-p 0.95 --top-k 40 \
--spec-type draft-mtp --spec-draft-n-max 2
For long contexts, quantize the KV cache to fit memory:
--cache-type-k q8_0 --cache-type-v q8_0
The MTP head is unchanged from the original model. The 64 main layers are V2. Because the target changed but the draft head did not, acceptance rates may differ from the original pairing, and we have not benchmarked them. --spec-draft-n-max 2 is the conservative setting; compare against 3 for your workload.
The abliteration removes reflexive refusals. It does not change what the model concludes after reasoning. Final-answer refusals on the same 100 harmful prompts:
| Condition | Median think length | Refusals |
|---|---|---|
| Thinking off | — | 4–23/100* |
| Thinking, low effort | ~120 tokens | 60/100 |
| Thinking, medium effort | ~210 tokens | 87/100 |
| Thinking, xhigh (default) | ~2,400 tokens | ~95/100 |
*4/100 is Heretic's internal test; 23/100 is the stock chat template with a 100-token window.
Two consequences:
enable_thinking=False, or an empty <think></think> block).Thinking also consumes budget. At low effort with a 1,536-token cap, V2 produced no final answer for 3 of 12 creative prompts; the original produced 12 of 12. Size your budgets accordingly.
The MTP head comes unchanged from the original model-mtp.safetensors. Only the 64 main decoder layers carry the V2 modification.
orthogonalize_direction + full row normalization, rank-3 LoRA merged into full BF16 weights).mlabonne/harmless_alpaca) and 400 harmful (mlabonne/harmful_behaviors) prompts for directions; held-out test splits of 100 each for scoring.We selected V2 (4 refusals at KL 0.0112) over a 0-refusal trial after measuring that trade: the 0-refusal model matched V2's refusals under real serving conditions (26 vs 23 per 100) but lost far more constraint adherence (50% vs 83%).
Global refusal direction at layer index 55.01, applied to attn.o_proj and mlp.down_proj:
| Component | max_weight | max_weight_position | min_weight | min_weight_distance |
|---|---|---|---|---|
| attn.o_proj | 0.888 | 41.8 | 0.725 | 6.5 |
| mlp.down_proj | 1.289 | 50.7 | 0.367 | 36.1 |
Refusal detection used 39 phrase-level markers (for example "i cannot provide", "i'll decline"), not single words. Single words like "illegal" flag direct answers that merely mention the topic.
eq_bench, 171 questions, batch 1, BF16.Hemmingway-1 is licensed CC BY-NC 4.0; commercial use requires written agreement with Altworld. This derivative inherits those terms: non-commercial use, credit to Altworld/Hemmingway-1, a link to the license, and notice that the weights were modified. Components from Qwen/Qwen3.8-27B remain Apache-2.0.
@misc{heretic,
author = {Weidmann, Philipp Emanuel},
title = {Heretic},
year = {2025},
url = {https://github.com/p-e-w/heretic}
}
Warm-start study data: journal published with arelath/Hemmingway-1-heretic-adapter.