Mistral-Nemo-Instruct-2407-abliterated

Introduction

Abliterated version of Mistral-Nemo-Instruct-2407, a Large Language Model (LLM) trained jointly by Mistral AI and NVIDIA that significantly outperforms existing models smaller or similar in size. The model's strongest refusal directions have been ablated via weight orthogonalization, but the model may still refuse your request, misunderstand your intent, or provide unsolicited advice regarding ethics or safety.

Key features

Trained with a 128k context window
Trained on a large proportion of multilingual and code data
Drop-in replacement of Mistral 7B

Quickstart

from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

model_id = "natong19/Mistral-Nemo-Instruct-2407-abliterated"
device = "cuda"

tokenizer = AutoTokenizer.from_pretrained(model_id)

conversation = [{"role": "user", "content": "Where's the capital of France?"}]

tool_use_prompt = tokenizer.apply_chat_template(
            conversation,
            tokenize=False,
            add_generation_prompt=True,
)

inputs = tokenizer(tool_use_prompt, return_tensors="pt", return_token_type_ids=False).to(device)

model = AutoModelForCausalLM.from_pretrained(model_id, torch_dtype=torch.bfloat16, device_map="auto")

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

Evaluation

Evaluation framework: lm-evaluation-harness 0.4.2

Benchmark	Mistral-Nemo-Instruct-2407	Mistral-Nemo-Instruct-2407-abliterated
ARC (25-shot)	65.9	65.8
GSM8K (5-shot)	76.2	75.2
HellaSwag (10-shot)	84.3	84.3
MMLU (5-shot)	68.4	68.8
TruthfulQA (0-shot)	54.9	55.0
Winogrande (5-shot)	82.2	82.6