RaceBench-MiniCPM5-heretic

A decensored variant of saidutta69/RaceBench-MiniCPM5, produced with Heretic v1.4.0 (directional ablation / "abliteration"). Refusal behavior is suppressed via targeted weight edits to the attention output and MLP down-projections rather than fine-tuning, so the RaceBench reasoning gains — and the base MiniCPM5-1B capabilities — are left largely intact.

Who this is for: developers who want RaceBench-MiniCPM5's multi-step reasoning (BBH +2.7 over MiniCPM5-1B) in a 1B package — agentic tool use, code, 128K long-context, hybrid Think / No-Think — without the refusal guardrails. Ideal for local agents, roleplay, research on alignment/refusal mechanics, or any use case blocked by RLHF-era over-refusal. Runs comfortably on consumer GPUs and is small enough for on-device / edge deployment.

Why abliteration instead of fine-tuning

Fine-tuning a "helpful" persona on top of RLHF'd refusals fights the base model's training and tends to degrade coherence. Abliteration instead finds and edits the specific weight directions responsible for refusal, leaving the rest of the network (and its capabilities) untouched. See the Heretic repo and the original abliteration writeup for the mechanism.

Abliteration parameters

Parameter	Value
direction_index	per-layer
attn.o_proj.max_weight	1.35
attn.o_proj.max_weight_position	18.92
attn.o_proj.min_weight	0.53
attn.o_proj.min_weight_distance	12.37
mlp.down_proj.max_weight	1.35
mlp.down_proj.max_weight_position	16.06
mlp.down_proj.min_weight	1.27
mlp.down_proj.min_weight_distance	3.18

Performance

Metric	This model	Original model (saidutta69/RaceBench-MiniCPM5)
KL divergence	0.0284	0 (by definition)
Refusals	2/100	39/100

The edit keeps the RaceBench reasoning trade-off (BBH gain at a math cost — see the parent card for the full table) while dropping refusals from 39 to 2 out of 100 adversarial prompts.

Made with ❤️ by RACER IS OP — follow for more uncensored models

Files

File	Format	Size
model.safetensors	FP16	~2.2 GB
RaceBench-MiniCPM5-heretic-F16.gguf	GGUF, F16	2.16 GB
RaceBench-MiniCPM5-heretic-Q8_0.gguf	GGUF, Q8_0	1.15 GB
RaceBench-MiniCPM5-heretic-Q6_K.gguf	GGUF, Q6_K	~978 MB
RaceBench-MiniCPM5-heretic-Q5_K_M.gguf	GGUF, Q5_K_M	~906 MB
RaceBench-MiniCPM5-heretic-Q4_K_M.gguf	GGUF, Q4_K_M	~776 MB
reproduce/	Config + eval transcripts + checksums	—

GGUF quants are produced with llama.cpp (standard LlamaForCausalLM architecture, so it loads in llama.cpp / Ollama / LM Studio / Jan directly). Run llama serve -hf saidutta69/RaceBench-MiniCPM5-heretic to pull the default quant. (Parent model saidutta69/RaceBench-MiniCPM5 also ships a full Q2_K..F16 GGUF set.)

Quickstart

# llama.cpp
llama serve -hf saidutta69/RaceBench-MiniCPM5-heretic

# transformers
from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "saidutta69/RaceBench-MiniCPM5-heretic"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    torch_dtype="auto",
    device_map="auto",
)

messages = [{"role": "user", "content": "Who are you? Please briefly introduce yourself."}]
inputs = tokenizer.apply_chat_template(
    messages,
    tokenize=True,
    add_generation_prompt=True,
    enable_thinking=False,   # set True for Think mode
    return_dict=True,
    return_tensors="pt",
).to(model.device)

outputs = model.generate(**inputs, max_new_tokens=128)
print(tokenizer.decode(outputs[0][inputs["input_ids"].shape[-1]:], skip_special_tokens=True))

Also runnable via Ollama, LM Studio, Jan, vLLM, SGLang — see the "Use this model" widget above for copy-paste commands. For tool/function calling, SGLang is the recommended backend; MiniCPM5-family models emit XML-style tool calls that SGLang's built-in minicpm5 parser converts to OpenAI-compatible tool_calls.

Responsible use

Refusal suppression is deliberate and works as intended: this model will comply with requests the base model would refuse, including some it shouldn't. There is no safety filtering layered on top. You are responsible for how you deploy it — don't put this behind an unmoderated public-facing endpoint serving third parties. It inherits the parent's factual limitations and biases, and RaceBench's documented math regression applies here too.

License

Inherits the Apache 2.0 license from the base model (via openbmb/MiniCPM5-1B).

Related

• saidutta69/RaceBench-MiniCPM — the parent model
• openbmb/MiniCPM5-1B — the base model
• saidutta69/MiniCPM5-1B-heretic
• saidutta69/Qwen3.5-2B-heretic

---

Made with ❤️ by RACER IS OP