LLaMA-Pro-8B Model Card
Model Description
LLaMA-Pro is a progressive version of the original LLaMA model, enhanced by the addition of Transformer blocks. It specializes in integrating both general language understanding and domain-specific knowledge, particularly in programming and mathematics.
Development and Training
Developed by Tencent's ARC Lab, LLaMA-Pro is an 8.3 billion parameter model. It's an expansion of LLaMA2-7B, further trained on code and math corpora totaling 80 billion tokens.
Intended Use
This model is designed for a wide range of NLP tasks, with a focus on programming, mathematics, and general language tasks. It suits scenarios requiring integration of natural and programming languages.
Performance
LLaMA-Pro demonstrates advanced performance across various benchmarks. It outperforms existing models in the LLaMA series in handling diverse tasks, showcasing its capability as an intelligent language agent.
Overall Performance on Languages, math and code tasks
Model | ARC | Hellaswag | MMLU | TruthfulQA | Winogrande | GSM8K | GSM8K-PoT | HumanEval | MBPP | Avg |
---|---|---|---|---|---|---|---|---|---|---|
LLAMA PRO (8B) | 54.10 | 77.94 | 47.88 | 39.04 | 73.95 | 17.89 | 25.42 | 28.66 | 33.20 | 44.2 |
LLaMA2-7B | 53.07 | 78.59 | 46.87 | 38.76 | 74.03 | 14.48 | 17.68 | 13.05 | 20.09 | 39.62 |
CodeLLaMA-7B | 39.93 | 60.80 | 31.12 | 37.82 | 64.01 | 5.16 | 25.20 | 33.50 | 41.40 | 37.66 |
LLAMA PRO-INSTRUCT | 52.30 | 76.88 | 52.57 | 48.80 | 72.53 | 43.59 | 55.61 | 44.51 | 37.88 | 53.8 |
Performance on GPT4 Evaluation
Model | MT Bench |
---|---|
Alpaca-13B | 4.53 |
CodeLLaMA-7B-Instruct | 5.71 |
Vicuna-7B | 6.17 |
LLaMA2-7B-Chat | 6.27 |
LLAMA PRO-INSTRUCT | 6.32 |
Limitations
While LLaMA-Pro addresses some limitations of previous models in the series, it may still encounter challenges specific to highly specialized domains or tasks.
Ethical Considerations
Users should be aware of potential biases in the model and use it responsibly, considering its impact on various applications.
- Downloads last month
- 1,158