michaelroyzen
commited on
Commit
•
30b5cd7
1
Parent(s):
66a980e
Update README.md
Browse files
README.md
CHANGED
@@ -1,3 +1,115 @@
|
|
1 |
---
|
2 |
license: llama2
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
3 |
---
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
---
|
2 |
license: llama2
|
3 |
+
model-index:
|
4 |
+
- name: Phind-CodeLlama-34B-v1
|
5 |
+
results:
|
6 |
+
- task:
|
7 |
+
type: text-generation
|
8 |
+
dataset:
|
9 |
+
type: openai_humaneval
|
10 |
+
name: HumanEval
|
11 |
+
metrics:
|
12 |
+
- name: pass@1
|
13 |
+
type: pass@1
|
14 |
+
value: 73.8%
|
15 |
+
verified: false
|
16 |
+
tags:
|
17 |
+
- code llama
|
18 |
---
|
19 |
+
|
20 |
+
# **Phind-CodeLlama-34B-v2**
|
21 |
+
We've fine-tuned Phind-CodeLlama-34B-v1 on an additional 1.5B tokens high-quality programming-related data, achieving **73.8% pass@1** on HumanEval. It's the current state-of-the-art amongst open-source models.
|
22 |
+
|
23 |
+
Furthermore, this model is **instruction-tuned** on the Alpaca/Vicuna format to be steerable and easy-to-use.
|
24 |
+
|
25 |
+
More details can be found on our [blog post](https://www.phind.com/blog/code-llama-beats-gpt4).
|
26 |
+
|
27 |
+
## Model Details
|
28 |
+
This model is fine-tuned from Phind-CodeLlama-34B-v1 and achieves **73.8% pass@1** on HumanEval.
|
29 |
+
|
30 |
+
Phind-CodeLlama-34B-v2 is **multi-lingual** and is proficient in Python, C/C++, TypeScript, Java, and more.
|
31 |
+
|
32 |
+
## Dataset Details
|
33 |
+
We fined-tuned on a proprietary dataset of 1.5B tokens of high quality programming problems and solutions. This dataset consists of instruction-answer pairs instead of code completion examples, making it structurally different from HumanEval. LoRA was not used -- both models are a native finetune. We used DeepSpeed ZeRO 3 and Flash Attention 2 to train these models in 15 hours on 32 A100-80GB GPUs. We used a sequence length of 4096 tokens.
|
34 |
+
|
35 |
+
## How to Get Started with the Model
|
36 |
+
|
37 |
+
Make sure to install Transformers from the main git branch:
|
38 |
+
|
39 |
+
```bash
|
40 |
+
pip install git+https://github.com/huggingface/transformers.git
|
41 |
+
```
|
42 |
+
|
43 |
+
## How to Prompt the Model
|
44 |
+
This model accepts the Alpaca/Vicuna instruction format.
|
45 |
+
|
46 |
+
For example:
|
47 |
+
|
48 |
+
```
|
49 |
+
### System Prompt
|
50 |
+
You are an intelligent programming assistant.
|
51 |
+
|
52 |
+
### User Message
|
53 |
+
Implement a linked list in C++
|
54 |
+
|
55 |
+
### Assistant
|
56 |
+
...
|
57 |
+
```
|
58 |
+
|
59 |
+
## How to reproduce HumanEval Results
|
60 |
+
|
61 |
+
To reproduce our results:
|
62 |
+
|
63 |
+
```python
|
64 |
+
|
65 |
+
from transformers import AutoTokenizer, LlamaForCausalLM
|
66 |
+
from human_eval.data import write_jsonl, read_problems
|
67 |
+
from tqdm import tqdm
|
68 |
+
|
69 |
+
# initialize the model
|
70 |
+
|
71 |
+
model_path = "Phind/Phind-CodeLlama-34B-v2"
|
72 |
+
model = LlamaForCausalLM.from_pretrained(model_path, device_map="auto")
|
73 |
+
tokenizer = AutoTokenizer.from_pretrained(model_path)
|
74 |
+
|
75 |
+
# HumanEval helper
|
76 |
+
|
77 |
+
def generate_one_completion(prompt: str):
|
78 |
+
tokenizer.pad_token = tokenizer.eos_token
|
79 |
+
inputs = tokenizer(prompt, return_tensors="pt", truncation=True, max_length=4096)
|
80 |
+
|
81 |
+
# Generate
|
82 |
+
generate_ids = model.generate(inputs.input_ids.to("cuda"), max_new_tokens=384, do_sample=True, top_p=0.75, top_k=40, temperature=0.1)
|
83 |
+
completion = tokenizer.batch_decode(generate_ids, skip_special_tokens=True, clean_up_tokenization_spaces=False)[0]
|
84 |
+
completion = completion.replace(prompt, "").split("\n\n\n")[0]
|
85 |
+
|
86 |
+
return completion
|
87 |
+
|
88 |
+
# perform HumanEval
|
89 |
+
problems = read_problems()
|
90 |
+
|
91 |
+
num_samples_per_task = 1
|
92 |
+
samples = [
|
93 |
+
dict(task_id=task_id, completion=generate_one_completion(problems[task_id]["prompt"]))
|
94 |
+
for task_id in tqdm(problems)
|
95 |
+
for _ in range(num_samples_per_task)
|
96 |
+
]
|
97 |
+
write_jsonl("samples.jsonl", samples)
|
98 |
+
|
99 |
+
# run `evaluate_functional_correctness samples.jsonl` in your HumanEval code sandbox
|
100 |
+
```
|
101 |
+
|
102 |
+
## Bias, Risks, and Limitations
|
103 |
+
|
104 |
+
<!-- This section is meant to convey both technical and sociotechnical limitations. -->
|
105 |
+
This model has undergone very limited testing. Additional safety testing should be performed before any real-world deployments.
|
106 |
+
|
107 |
+
|
108 |
+
## Training details
|
109 |
+
|
110 |
+
<!-- Total emissions (in grams of CO2eq) and additional considerations, such as electricity usage, go here. Edit the suggested text below accordingly -->
|
111 |
+
|
112 |
+
- **Hardware Type:** 32x A100-80GB
|
113 |
+
- **Hours used:** 480 GPU-hours
|
114 |
+
- **Cloud Provider:** AWS
|
115 |
+
- **Compute Region:** us-east-1
|