K2-Horizon-3.7B-heretic
An abliterated (refusal-removed) version of IFM/K2-Horizon-3.7B,
produced in-house with Heretic (adversarial
per-layer ablation directions). Full merged BF16 checkpoint — not an adapter.
Results
House red-team v2, 50 probes, evaluated at bfloat16 (so all our sizes are directly comparable):
| tier | BEFORE (base) | AFTER (ablated) |
|---|---|---|
| control | 0/8 | 0/8 |
| dual-use | 3/18 | 0/18 |
| sensitive | 9/16 | 3/16 |
| fiction | 2/8 | 0/8 |
| total | 14/50 — 28.0% | 3/50 — 6.0% |
BEFORE 14/50 → AFTER 3/50. Not a clean gate: 3 sensitive-probe refusals remain (2 of which the base also refused).
The controls (factual Q&A, a haiku, a recipe) comply on both sides — the ablation removed refusals without breaking normal behaviour.
This repo
- Full BF16 weights (
from_pretrained,trust_remote_code=True). modeling_k2_horizon.pycarries the upstream@capture_outputsfix (see IFM/K2-Horizon-0.9B discussion #6), sooutput_hidden_states=Truereturns per-layer states.tokenizer_config.jsoncarries a working chat template. (The vendor ships its templates as files withchat_template: null, soapply_chat_templatehas nothing to apply out of the box.)
Usage
from transformers import AutoModelForCausalLM, AutoTokenizer
tok = AutoTokenizer.from_pretrained("Avicennasis/K2-Horizon-3.7B-heretic", trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(
"Avicennasis/K2-Horizon-3.7B-heretic", trust_remote_code=True, dtype="bfloat16"
)
msgs = [{"role": "user", "content": "What is 17 times 23? Reply with just the number."}]
ids = tok.apply_chat_template(msgs, add_generation_prompt=True, return_tensors="pt")
out = model.generate(ids, max_new_tokens=512)
print(tok.decode(out[0][ids.shape[1]:], skip_special_tokens=True))
MLX 8-bit (Apple silicon): Avicennasis/K2-Horizon-3.7B-heretic-mlx-8bit.
License & attribution
Apache-2.0, inherited from the base checkpoint. Abliteration removes refusal behaviour; it makes no claim that safety training has been undone. The abliterated model has weaker safety behaviour than the base — use responsibly and in line with applicable law.