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.

Open LLM Leaderboard Evaluation Results

Detailed results can be found here

Metric	Value
Avg.	34.72
IFEval (0-Shot)	77.61
BBH (3-Shot)	49.07
MATH Lvl 5 (4-Shot)	21.22
GPQA (0-shot)	6.15
MuSR (0-shot)	12.43
MMLU-PRO (5-shot)	41.83