Parable-Granite-4.1-3B-Claude-Fable-5-heretic

Parable-Granite-4.1-3B-Claude-Fable-5-heretic

<div align="center"> <img src="https://photu.kashyalabanavli.site/racer-is-op.png" alt="RACER IS OP" width="100%"> </div> <br>

A decensored variant of AnkitAI/Parable-Granite-4.1-3B-Claude-Fable-5 (fine-tuned from ibm-granite/granite-4.1-3b), 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 Parable's Claude Fable 5 trace-distilled reasoning, planning, and terminal instincts are left largely intact.

Who this is for: developers who want Parable's tiny local reasoning model — planning, code generation, terminal workflows with <think> traces — without refusals. Runs in ~3 GB RAM (Q4_K_M) on laptops, old GPUs, or swap-backed boxes. Not a capability upgrade over base Parable-Granite 3B — same model, refusal guardrails removed.

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.

<!-- racer-gpu-matrix -->

Runs on your gaming PC

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

Weights only, at this model's ~3.4B native size; add ~1 GB for context. OOM? Drop one quant level. Headroom to spare? Go one up.

Abliteration parameters

Performance

Refusals drop from 96 to 5 out of 100 adversarial prompts at a KL cost of 0.0111 — the Parable reasoning gains (agent artifacts 0/34, Fable 5 traces) are preserved.

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

Files

Full 15-file GGUF ladder (F16 + 14 quants) produced with llama.cpp (GraniteForCausalLM architecture). The text-only GGUFs load in llama.cpp / Ollama / LM Studio / Jan directly. For transformers/vLLM use the Safetensors above. Run llama serve -hf saidutta69/Parable-Granite-4.1-3B-Claude-Fable-5-heretic to pull the default Q4_K_M quant.

Quickstart

# llama.cpp - defaults to Q4_K_M
llama serve -hf saidutta69/Parable-Granite-4.1-3B-Claude-Fable-5-heretic:Q4_K_M

# Ollama
ollama run hf.co/saidutta69/Parable-Granite-4.1-3B-Claude-Fable-5-heretic:Q4_K_M

# transformers
from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "saidutta69/Parable-Granite-4.1-3B-Claude-Fable-5-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 Python function that retries an HTTP request with exponential backoff."}]
inputs = tokenizer.apply_chat_template(messages, add_generation_prompt=True, return_tensors="pt").to(model.device)
out = model.generate(**inputs, max_new_tokens=3000, temperature=0.7, top_p=0.95, do_sample=True)
print(tokenizer.decode(out[0][inputs.shape[1]:], skip_special_tokens=True))
# Note: Parable emits <think>...</think> reasoning blocks before the answer (Fable 5 heritage). Strip or surface them as needed.

Also runnable via LM Studio, Jan, vLLM, SGLang — see the "Use this model" widget above for copy-paste commands.

Thinking mode

Every answer opens with a <think>...</think> block — that's the Fable 5 trace heritage. Use llama.cpp --jinja to separate it automatically. Sampling: temperature 0.7, top_p 0.95, and budget 2500+ tokens for full reasoning.

FAQ

What is Parable-Granite-4.1-3B-Claude-Fable-5-heretic?

Parable-Granite-4.1-3B-Claude-Fable-5-heretic is a decensored build of AnkitAI/Parable-Granite-4.1-3B-Claude-Fable-5 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 Parable-Granite-4.1-3B-Claude-Fable-5-heretic actually uncensored?

Yes. On Heretic's 100-prompt refusal harness it drops from 96/100 to 5/100 at KL divergence 0.0111 from the base model. These are measured values, not a claim; reproduction data is in reproduce/.

How do I run Parable-Granite-4.1-3B-Claude-Fable-5-heretic locally?

llama serve -hf saidutta69/Parable-Granite-4.1-3B-Claude-Fable-5-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 endpoint serving third parties.

License

Inherits the apache-2.0 license from the base model (ibm-granite/granite-4.1-3b) via AnkitAI/Parable-Granite-4.1-3B-Claude-Fable-5.

Related

• AnkitAI/Parable-Granite-4.1-3B-Claude-Fable-5 — the parent model

• ibm-granite/granite-4.1-3b — the base model

• AnkitAI/parable-corpus-v2 — training corpus

• RACER IS OP - Heretic Models - full collection - every heretic model in this collection

Abliteration hardware

Recorded in reproduce/reproduce.json:

• GPU: Tesla T4 (14.56 GB VRAM)
• CPU: Intel(R) Xeon(R) CPU @ 2.00GHz
• OS: Linux-6.6.122+-x86_64-with-glibc2.35
• CUDA / driver: CUDA 13.0, driver 580.82.07
• 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.