metadata
library_name: transformers
tags:
- Chat Model
- SFT
- RLHF
license: llama3
pipeline_tag: text-generation
Llama3-PBM-Nova-70B
Introduction
Llama3-PBM-Nova-70B is a chat model developed by PKU-Baichuan-MLSysLab, based on the Llama3-70B. In order to better utilize open-source data, we've performed deduplication, quality filtering, and data synthesis on it. Then, through Supervised Fine-Tuning (SFT) and Reinforcement Learning from Human Feedback (RLHF), we've significantly enhanced the base model's performance.
- Paper: Baichuan Alignment Technical Report
- Developed by: PKU-Baichuan-MLSysLab
- Base Model: Llama-3-70B
- Model Type: Chat Model
- Training Method: SFT + RLHF
- Release Date: August 2024
Evaluation
| Model | Arena-Hard | MixEval-Hard | Alpaca-Eval 2.0 |
|---|---|---|---|
| GPT-4Turbo (04/09) | 82.6% | 62.6 | 55.0% |
| GPT-4o (05/13) | 79.2% | 64.7 | 57.5% |
| Gemini 1.5 Pro | 72.0% | 58.3 | - |
| Llama3-PBM-Nova-70B | 74.5% | 58.1 | 56.9% |
| Llama-3.1-70B-Instruct | 55.7% | 61.25 | 38.1% |
| Llama-3-70B-Instruct | 46.6% | 55.9 | 34.4% |
Usage
Below is an example of how to use this model based on the Transformers library.
import transformers
import torch
model_id = "PKU-Baichuan-MLSystemLab/Llama3-PBM-Nova-70B"
pipeline = transformers.pipeline(
"text-generation",
model=model_id,
model_kwargs={"torch_dtype": torch.bfloat16},
device_map="auto",
)
messages = [
{"role": "user", "content": "Who are you?"},
]
terminators = [
pipeline.tokenizer.eos_token_id,
pipeline.tokenizer.convert_tokens_to_ids("<|eot_id|>")
]
outputs = pipeline(
messages,
max_new_tokens=256,
eos_token_id=terminators,
do_sample=True,
temperature=0.6,
top_p=0.9,
)
print(outputs[0]["generated_text"][-1])