Orpo-Llama-3.2-1B-15k

AdamLucek/Orpo-Llama-3.2-1B-15k is an ORPO fine tuned version of meta-llama/Llama-3.2-1B on a subset of 15,000 shuffled entries of mlabonne/orpo-dpo-mix-40k.

Trained for 7 hours on an L4 GPU with this training script, modified from Maxime Labonne's original guide

For full model details, refer to the base model page meta-llama/Llama-3.2-1B

Evaluations

In comparsion to AdamLucek/Orpo-Llama-3.2-1B-40k using lm-evaluation-harness.

Benchmark	15k Accuracy	15k Normalized	40k Accuracy	40k Normalized	Notes
AGIEval	22.14%	21.01%	23.57%	23.26%	0-Shot Average across multiple reasoning tasks
GPT4ALL	51.15%	54.38%	51.63%	55.00%	0-Shot Average across all categories
TruthfulQA	42.79%	N/A	42.14%	N/A	MC2 accuracy
MMLU	31.22%	N/A	31.01%	N/A	5-Shot Average across all categories
Winogrande	61.72%	N/A	61.12%	N/A	0-shot evaluation
ARC Challenge	32.94%	36.01%	33.36%	37.63%	0-shot evaluation
ARC Easy	64.52%	60.40%	65.91%	60.90%	0-shot evaluation
BoolQ	50.24%	N/A	52.29%	N/A	0-shot evaluation
PIQA	75.46%	74.37%	75.63%	75.19%	0-shot evaluation
HellaSwag	48.56%	64.71%	48.46%	64.50%	0-shot evaluation

Using this Model

from transformers import AutoTokenizer
import transformers
import torch

# Load Model and Pipeline
model = "AdamLucek/Orpo-Llama-3.2-1B-15k"

pipeline = transformers.pipeline(
    "text-generation",
    model=model,
    torch_dtype=torch.float16,
    device_map="auto",
)

# Load Tokenizer
tokenizer = AutoTokenizer.from_pretrained(model)

# Generate Message
messages = [{"role": "user", "content": "What is a language model?"}]
prompt = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
outputs = pipeline(prompt, max_new_tokens=256, do_sample=True, temperature=0.7, top_k=50, top_p=0.95)
print(outputs[0]["generated_text"])

Training Statistics