Smaug-Llama-3-70B-Instruct
Built with Meta Llama 3
This model was built using a new Smaug recipe for improving performance on real world multi-turn conversations applied to meta-llama/Meta-Llama-3-70B-Instruct.
The model outperforms Llama-3-70B-Instruct substantially, and is on par with GPT-4-Turbo, on MT-Bench (see below).
EDIT: Smaug-Llama-3-70B-Instruct is the top open source model on Arena-Hard currently! It is also nearly on par with Claude Opus - see below.
We are conducting additional benchmark evaluations and will add those when available.
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
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 | 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 |
MT-Bench
########## First turn ##########
score
model turn
Smaug-Llama-3-70B-Instruct 1 9.40000
GPT-4-Turbo 1 9.37500
Meta-Llama-3-70B-Instruct 1 9.21250
########## Second turn ##########
score
model turn
Smaug-Llama-3-70B-Instruct 2 9.0125
GPT-4-Turbo 2 9.0000
Meta-Llama-3-70B-Instruct 2 8.8000
########## Average ##########
score
model
Smaug-Llama-3-70B-Instruct 9.206250
GPT-4-Turbo 9.187500
Meta-Llama-3-70B-Instruct 9.006250
Model | First turn | Second Turn | Average |
---|---|---|---|
Smaug-Llama-3-70B-Instruct | 9.40 | 9.01 | 9.21 |
GPT-4-Turbo | 9.38 | 9.00 | 9.19 |
Meta-Llama-3-70B-Instruct | 9.21 | 8.80 | 9.01 |
OpenLLM Leaderboard Manual Evaluation
Model | ARC | Hellaswag | MMLU | TruthfulQA | Winogrande | GSM8K* | Average |
---|---|---|---|---|---|---|---|
Smaug-Llama-3-70B-Instruct | 70.6 | 86.1 | 79.2 | 62.5 | 83.5 | 90.5 | 78.7 |
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.
This version of Smaug uses new techniques and new data compared to Smaug-72B, and more information will be released later on. For now, see the previous Smaug paper: https://arxiv.org/abs/2402.13228.
- Downloads last month
- 3,611