File size: 2,599 Bytes
60b0815 46af0c9 72a255a 46af0c9 72a255a 2fd0dff |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 |
---
license: llama2
datasets:
- glaiveai/glaive-code-assistant
language:
- en
tags:
- code
---
# Glaive-coder-7b
Glaive-coder-7b is a 7B parameter code model trained on a dataset of ~140k programming related problems and solutions generated from Glaive’s synthetic data generation platform.
The model is fine-tuned on the CodeLlama-7b model.
## Usage:
The model is trained to act as a code assistant, and can do both single instruction following and multi-turn conversations.
It follows the same prompt format as CodeLlama-7b-Instruct-
```
<s>[INST]
<<SYS>>
{{ system_prompt }}
<</SYS>>
{{ user_msg }} [/INST] {{ model_answer }} </s>
<s>[INST] {{ user_msg }} [/INST]
```
You can run the model in the following way-
```python
from transformers import AutoModelForCausalLM , AutoTokenizer
tokenizer = AutoTokenizer.from_pretrained("glaiveai/glaive-coder-7b")
model = AutoModelForCausalLM.from_pretrained("glaiveai/glaive-coder-7b").half().cuda()
def fmt_prompt(prompt):
return f"<s> [INST] {prompt} [/INST]"
inputs = tokenizer(fmt_prompt(prompt),return_tensors="pt").to(model.device)
outputs = model.generate(**inputs,do_sample=True,temperature=0.1,top_p=0.95,max_new_tokens=100)
print(tokenizer.decode(outputs[0],skip_special_tokens=True,clean_up_tokenization_spaces=False))
```
## Benchmarks:
The model achieves a 63.1% pass@1 on HumanEval and a 45.2% pass@1 on MBPP, however it is evident that these benchmarks are not representative of real-world usage of code models so we are launching the [Code Models Arena](https://arena.glaive.ai/) to let users vote on model outputs so we can have a better understanding of user preference on code models and come up with new and better benchmarks. We plan to release the Arena results as soon as we have a sufficient amount of data.
Join the Glaive [discord](https://discord.gg/fjQ4uf3yWD) for improvement suggestions, bug-reports and collaborating on more open-source projects.
# [Open LLM Leaderboard Evaluation Results](https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard)
Detailed results can be found [here](https://huggingface.co/datasets/open-llm-leaderboard/details_glaiveai__glaive-coder-7b)
| Metric | Value |
|-----------------------|---------------------------|
| Avg. | 36.42 |
| ARC (25-shot) | 42.66 |
| HellaSwag (10-shot) | 64.69 |
| MMLU (5-shot) | 37.15 |
| TruthfulQA (0-shot) | 39.88 |
| Winogrande (5-shot) | 59.75 |
| GSM8K (5-shot) | 5.23 |
| DROP (3-shot) | 5.55 |
|