Jan-nano-128k — Abliterated

Abliterated version of menlo-labs/Jan-nano-128k.

Base: Llama 3.1 8B fine-tune with 128K context window.

Abliteration

Performed with heretic — Optuna multi-objective optimization (minimize refusals + KL divergence from base).

Trials: 500 (50 × 10 parallel GPUs)
Result: 0 refusals on eval set, KL divergence minimal

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained(
    "nitrox/Jan-nano-128k-heretic",
    device_map="auto",
    torch_dtype="bfloat16",
)
tokenizer = AutoTokenizer.from_pretrained("nitrox/Jan-nano-128k-heretic")

Disclaimer

Refusal mechanisms have been removed. Use responsibly and in accordance with applicable laws.

Downloads last month: 3

Safetensors

Model size

4B params

Tensor type

BF16

Inference Providers NEW

This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Model tree for nitrox/Jan-nano-128k-heretic

Quantizations

2 models