Dracarys2-Llama-3.1-70B-Instruct

Built with Meta Llama 3

Introduction

We introduce the latest in the Smaug series, the Dracarys family of finetunes targeting coding performance improvements across a variety of base models.

This variant is a finetune of meta-llama/Meta-Llama-3.1-70B-Instruct

Compared to meta-llama/Meta-Llama-3.1-70B-Instruct, Dracarys has better LiveCodeBench scores (see evaluation results below).

Model Description

Developed by: Abacus.AI
License: https://llama.meta.com/llama3/license/
Finetuned from model: meta-llama/Meta-Llama-3.1-70B-Instruct.

How to use

The prompt format is unchanged from Llama 3 70B Instruct (see evaluations for prompt details for LCB)

Use with transformers

See the snippet below for usage with Transformers:

import transformers
import torch

model_id = "abacusai/Dracarys-72B-Instruct"

pipeline = transformers.pipeline(
    "text-generation",
    model=model_id,
    model_kwargs={"torch_dtype": torch.bfloat16},
    device_map="auto",
)

messages = [
    {"role": "system", "content": "You are data science coding assistant that generates Python code using Pandas and Numpy."},
    {"role": "user", "content": "Write code to select rows from the dataframe `df` having the maximum `temp` for each `city`"},
]

prompt = pipeline.tokenizer.apply_chat_template(
        messages, 
        tokenize=False, 
        add_generation_prompt=True
)

terminators = [
    pipeline.tokenizer.eos_token_id,
    pipeline.tokenizer.convert_tokens_to_ids("<|eot_id|>"),
    pipeline.tokenizer.convert_tokens_to_ids("<|end_of_text|>"),
]

outputs = pipeline(
    prompt,
    max_new_tokens=256,
    eos_token_id=terminators,
    do_sample=True,
    temperature=0.6,
    top_p=0.9,
)
print(outputs[0]["generated_text"][len(prompt):])

Evaluation Results

LiveCodeBench

Model	Code Generation	Code Execution	Test Output Prediction
Dracarys2-Llama-3.1-70B-Instruct	33.44	48.26	52.10
Meta-Llama-3.1-70B-Instruct	32.23	48.768	41.40

Breakdown of LiveCodeBench CodeGeneration

Model	Easy	Medium	Hard
Dracarys2-Llama-3.1-70B-Instruct	71.29	18.48	3.57
Meta-Llama-3.1-70B-Instruct	68.4	17.99	3.57

Breakdown of LiveCodeBench CodeExecution

Model	COT	Non-COT
Dracarys2-Llama-3.1-70B-Instruct	75.55	48.26
Meta-Llama-3.1-70B-Instruct	70.14	48.768

Breakdown of LiveCodeBench TestOutputPrediction

Model	Easy	Medium	Hard
Dracarys2-Llama-3.1-70B-Instruct	63.53	47.30	43.61
Meta-Llama-3.1-70B-Instruct	51.22	35.91	34.30

LiveBench(Aug update)

Model	Global Average	Coding Average	Reasoning Average	Mathematics Average	Data Analysis Average	Language Average	IF Average
Dracarys2-Llama-3.1-70B-Instruct	47.8	36.3	47.3	38.9	46.1	41.5	76.6
Meta-Llama-3.1-70B-Instruct	45.1	30.7	35.3	37.0	48.4	42.1	77.2