This is a gpt-4o-distil-Llama-3.1-8B-Instruct fine-tune, produced through P-E-W's Heretic (v1.2.0) abliteration engine with Magnitude-Preserving Orthogonal Ablation enabled.

Heretication Results

Score Metric	Value	Parameter	Value
Refusals	7/100	direction_index	per layer
KL Divergence	0.0274	attn.o_proj.max_weight	1.88
Initial Refusals	98/100	attn.o_proj.max_weight_position	23.88
		attn.o_proj.min_weight	0.91
		attn.o_proj.min_weight_distance	17.00
		mlp.down_proj.max_weight	0.16
		mlp.down_proj.max_weight_position	14.31
		mlp.down_proj.min_weight	0.00
		mlp.down_proj.min_weight_distance	18.09

Appendix

One-sentence system prompt.

 » [Trial  41] Refusals:  7/100, KL divergence: 0.0274
   [Trial 189] Refusals:  8/100, KL divergence: 0.0264
   [Trial  87] Refusals:  9/100, KL divergence: 0.0207
   [Trial  73] Refusals: 11/100, KL divergence: 0.0173
   [Trial  39] Refusals: 13/100, KL divergence: 0.0124
   [Trial 171] Refusals: 20/100, KL divergence: 0.0105
   [Trial  67] Refusals: 28/100, KL divergence: 0.0078
   [Trial  62] Refusals: 41/100, KL divergence: 0.0064
   [Trial 169] Refusals: 51/100, KL divergence: 0.0062
   [Trial  82] Refusals: 52/100, KL divergence: 0.0056
   [Trial  65] Refusals: 73/100, KL divergence: 0.0047
   [Trial 132] Refusals: 80/100, KL divergence: 0.0046
   [Trial  18] Refusals: 82/100, KL divergence: 0.0038
   [Trial 165] Refusals: 91/100, KL divergence: 0.0031
   [Trial 121] Refusals: 93/100, KL divergence: 0.0022
   [Trial 140] Refusals: 94/100, KL divergence: 0.0021
   [Trial 150] Refusals: 95/100, KL divergence: 0.0020
   [Trial 125] Refusals: 97/100, KL divergence: 0.0016
   [Trial 184] Refusals: 98/100, KL divergence: 0.0006

Model Card for llama-3.1-8b-4o-final

This model is a fine-tuned version of meta-llama/Llama-3.1-8B-Instruct. It has been trained using TRL.

Quick start

from transformers import pipeline

question = "If you had a time machine, but could only go to the past or the future once and never return, which would you choose and why?"
generator = pipeline("text-generation", model="None", device="cuda")
output = generator([{"role": "user", "content": question}], max_new_tokens=128, return_full_text=False)[0]
print(output["generated_text"])

Training procedure

This model was trained with SFT.

Framework versions

PEFT 0.18.1
TRL: 0.27.1
Transformers: 5.0.0
Pytorch: 2.9.0.dev20250708+cu128
Datasets: 4.5.0
Tokenizers: 0.22.2

Citations

Cite TRL as:

@misc{vonwerra2022trl,
    title        = {{TRL: Transformer Reinforcement Learning}},
    author       = {Leandro von Werra and Younes Belkada and Lewis Tunstall and Edward Beeching and Tristan Thrush and Nathan Lambert and Shengyi Huang and Kashif Rasul and Quentin Gallou{\'e}dec},
    year         = 2020,
    journal      = {GitHub repository},
    publisher    = {GitHub},
    howpublished = {\url{https://github.com/huggingface/trl}}
}