Dracarys-72B-Instruct

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 Qwen2-72B-Instruct

Compared to Qwen2-72B-Instruct, Dracarys has better LiveCodeBench scores (see evaluation results below).

Model Description

• Developed by: Abacus.AI
• License: https://huggingface.co/Qwen/Qwen2-72B-Instruct/blob/main/LICENSE
• Finetuned from model: Qwen2-72B-Instruct.

How to use

The prompt format is unchanged from Qwen2-72B-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|>")
]

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
Dracarys-72B-Instruct	33.57	62.96	58.93
Qwen2-72B-Instruct	32.92	58.95	55.88

Breakdown of LiveCodeBench CodeGeneration

Model	Easy	Medium	Hard
Dracarys-72B-Instruct	64.16	25.06	3.64
Qwen2-72B-Instruct	65.83	22.28	3.11

Breakdown of LiveCodeBench TestOutputPrediction

Model	Easy	Medium	Hard
Dracarys-72B-Instruct	65.37	58.74	46.38
Qwen2-72B-Instruct	63.19	54.08	46.52

LiveBench (July update)

Model	Global Average	Coding Average	Language Average	Mathematics Average	Data Analysis Average	Reasoning Average	IF Average
Dracarys-72B-Instruct	41.20	38.95	31.17	42.77	26.24	40	68.08
Qwen2-72B-Instruct	40.15	32.38	29.21	43.44	26.24	41.33	68.27