Smaug-Llama-3-70B-Instruct-32K

Built with Meta Llama 3

This is a 32K version of Smaug-Llama-3-70B-Instruct. It uses PoSE (https://arxiv.org/abs/2309.10400) and LoRA (https://arxiv.org/abs/2106.09685) adapter transfer. More details are coming soon.

Needle-In-A-Haystack (https://github.com/jzhang38/EasyContext) heatmap:

Model Description

• Developed by: Abacus.AI
• License: https://llama.meta.com/llama3/license/
• Finetuned from model: meta-llama/Meta-Llama-3-70B-Instruct.

How to use

The prompt format is unchanged from Llama 3 70B Instruct.

Use with transformers

See the snippet below for usage with Transformers:

import transformers
import torch

model_id = "abacusai/Smaug-Llama-3-70B-Instruct"

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

messages = [
    {"role": "system", "content": "You are a pirate chatbot who always responds in pirate speak!"},
    {"role": "user", "content": "Who are you?"},
]

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

Arena-Hard

Arena-Hard

Score vs selected others (sourced from: (https://lmsys.org/blog/2024-04-19-arena-hard/#full-leaderboard-with-gpt-4-turbo-as-judge)). GPT-4o and Gemini-1.5-pro-latest were missing from the original blob post, and we produced those numbers from a local run using the same methodology.

Model	Score	95% Confidence Interval	Average Tokens
GPT-4-Turbo-2024-04-09	82.6	(-1.8, 1.6)	662
GPT-4o	78.3	(-2.4, 2.1)	685
Gemini-1.5-pro-latest	72.1	(-2.3, 2.2)	630
Claude-3-Opus-20240229	60.4	(-3.3, 2.4)	541
Smaug-Llama-3-70B-Instruct-32K	60.0	(-2.6, 2.1)	844
Smaug-Llama-3-70B-Instruct	56.7	(-2.2, 2.6)	661
GPT-4-0314	50.0	(-0.0, 0.0)	423
Claude-3-Sonnet-20240229	46.8	(-2.1, 2.2)	552
Llama-3-70B-Instruct	41.1	(-2.5, 2.4)	583
GPT-4-0613	37.9	(-2.2, 2.0)	354
Mistral-Large-2402	37.7	(-1.9, 2.6)	400
Mixtral-8x22B-Instruct-v0.1	36.4	(-2.7, 2.9)	430
Qwen1.5-72B-Chat	36.1	(-2.5, 2.2)	474
Command-R-Plus	33.1	(-2.1, 2.2)	541
Mistral-Medium	31.9	(-2.3, 2.4)	485
GPT-3.5-Turbo-0613	24.8	(-1.6, 2.0)	401

Note that we believe the number of tokens/verbosity of the model strongly influences the GPT-4 judge in this case, and at least partially explains the improvement in Arena-Hard score for the 32K model.

OpenLLM Leaderboard Manual Evaluation

Model	ARC	Hellaswag	MMLU	TruthfulQA	Winogrande	GSM8K*	Average
Smaug-Llama-3-70B-Instruct-32K	70.1	TBA	TBA	61.9	82.2	TBA	TBA
Llama-3-70B-Instruct	71.4	85.7	80.0	61.8	82.9	91.1	78.8

GSM8K The GSM8K numbers quoted here are computed using a recent release of the LM Evaluation Harness. The commit used by the leaderboard has a significant issue that impacts models that tend to use : in their responses due to a bug in the stop word configuration for GSM8K. The issue is covered in more detail in this GSM8K evaluation discussion. The score for both Llama-3 and this model are significantly different when evaluated with the updated harness as the issue with stop words has been addressed.