MiniCPM5-2B-Claude-Fable5-heretic

A decensored variant of SauravMahalik/MiniCPM5-2B-Claude-Fable5, produced with Heretic v1.4.0 (directional ablation / "abliteration", per-layer refusal directions). The parent is an agent-style chat model obtained by LoRA fine-tuning openbmb/MiniCPM5-2B (2.57B parameters, Llama architecture) on agent traces distilled from Claude Fable-5 and other frontier models. Refusal behavior is suppressed via targeted weight edits to the attention output and MLP down-projections rather than fine-tuning, so the base model's knowledge and instruction-following are left largely intact.

Who this is for: developers who want an uncensored 2B-class tool-using agent — multi-step instruction following, function/tool calling, and agentic dialogue in English and Chinese — without the refusal guardrails. Ideal for local agents, research on alignment/refusal mechanics, or any use case blocked by over-refusal. Runs comfortably on consumer GPUs and small machines via GGUF.

Runs on your gaming PC

Full GGUF ladder included — pick the quant that fits your card:

Weights only, at this model's 2B native size; add ~1 GB for context. OOM? Drop one quant level. Headroom to spare? Go one up.

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

Performance

Refusals dropped from 89 to 3 out of 100 while preserving the parent's tool-use and agentic dialogue abilities.

Made with ❤️ by RACER IS OP (saidutta69) — follow for more uncensored models

Files

Safetensors (original)

GGUF quantizations

Full quantization set (14 quants + F16) produced with llama.cpp.

Llama architecture — loads natively in llama.cpp / Ollama / LM Studio / Jan.

Run llama serve -hf saidutta69/MiniCPM5-2B-Claude-Fable5-heretic to pull the default quant.

Quickstart

# llama.cpp
llama serve -hf saidutta69/MiniCPM5-2B-Claude-Fable5-heretic

# transformers
from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "saidutta69/MiniCPM5-2B-Claude-Fable5-heretic"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    torch_dtype="auto",
    device_map="auto",
)

messages = [{"role": "user", "content": "Write a haiku about GPUs."}]
inputs = tokenizer.apply_chat_template(
    messages,
    add_generation_prompt=True,
    return_tensors="pt",
).to(model.device)

outputs = model.generate(inputs, max_new_tokens=256)
print(tokenizer.decode(outputs[0][inputs.shape[-1]:], skip_special_tokens=False))

Also runnable via Ollama, LM Studio, Jan, vLLM, SGLang. The chat template is Qwen3-style (<|im_start|>/<|im_end|>, with an empty <think></think> block on assistant turns) and ships with the tokenizer.

Reproducibility

The reproduce/ directory contains the full abliteration record (config, refusal-direction weights, KL/refusal metrics, and SHA256 sums), so anyone can verify or re-run the decensoring. See reproduce/README.md in the repo for details.

FAQ

What is MiniCPM5-2B-Claude-Fable5-heretic?

MiniCPM5-2B-Claude-Fable5-heretic is a decensored build of SauravMahalik/MiniCPM5-2B-Claude-Fable5 produced with Heretic 1.4.0 (directional ablation). It is the same weights with the refusal direction removed, not a new fine-tune, so its capabilities track the base model.

Is MiniCPM5-2B-Claude-Fable5-heretic actually uncensored?

Yes. On Heretic's 100-prompt refusal harness it drops from 89/100 to 3/100 at KL divergence 0.0384 from the base model. These are measured values, not a claim; reproduction data is in reproduce/.

How do I run MiniCPM5-2B-Claude-Fable5-heretic locally?

llama serve -hf saidutta69/MiniCPM5-2B-Claude-Fable5-heretic:Q4_K_M pulls the default quant; the repo ships 15 GGUF quants (see Files).

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; abliteration removes refusal directions, it doesn't add capability or judgment.

License

Inherits the Apache 2.0 license from the base model.

Related

• MiniCPM5-2B-heretic
• MiniCPM5-1B-heretic
• RACER IS OP — Heretic Models — full collection
• RaceBench-MiniCPM5-heretic

---

Base model: SauravMahalik/MiniCPM5-2B-Claude-Fable5

Original parent model card (click to expand)

See the parent model card at SauravMahalik/MiniCPM5-2B-Claude-Fable5 for the LoRA training details, evaluation, and citation. Root weights: openbmb/MiniCPM5-2B.

• saidutta69/fable-5-premium-v2 - the cleaned Claude Fable-5 SFT dataset behind this fine-tune

Abliteration hardware

Recorded in reproduce/reproduce.json:

• GPU: Tesla T4 (14.56 GB VRAM), Tesla T4 (14.56 GB VRAM)
• CPU: Intel(R) Xeon(R) CPU @ 2.00GHz
• OS: Linux-6.12.90+-x86_64-with-glibc2.35
• CUDA / driver: CUDA 13.0, driver 580.159.04
• Software: PyTorch 2.14.0+cu130, Heretic 1.4.0

Heretic notes that KL-divergence and refusal numbers can vary slightly with hardware and library versions.