michaelroyzen commited on
Commit
6025980
1 Parent(s): a5d9b2a

Update README.md

Browse files
Files changed (1) hide show
  1. README.md +87 -0
README.md CHANGED
@@ -1,3 +1,90 @@
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: 67.6%
15
+ verified: false
16
  ---
17
+
18
+ # **Phind-CodeLlama-34B-v1**
19
+ We've fine-tuned CodeLlama-34B and CodeLlama-34B-Python on an internal Phind dataset that achieve 67.6% and 69.5% pass@1 on HumanEval, respectively. GPT-4 achieves 67%. We've applied OpenAI's decontamination methodology to our dataset to ensure result validity.
20
+
21
+ ## Model Details
22
+ This model is fine-tuned from CodeLlama-34B and achieves 67.6% pass@1 on HumanEval.
23
+
24
+ ## Dataset Details
25
+ We fined-tuned on a proprietary dataset of ~80k high quality programming problems and solutions. This dataset consists of instruction-answer pairs instead of code completion examples, making it structurally different from HumanEval. The Phind models were trained for 2 epochs, for a total of ~160k examples shown. LoRA was not used -- both models are a native finetune. We used DeepSpeed ZeRO 3 and Flash Attention 2 to train these models in three hours on 32 A100-80GB GPUs. We used a sequence length of 4096 tokens.
26
+
27
+ ## How to Get Started with the Model
28
+
29
+ Make sure to install Transformers from the main git branch:
30
+
31
+ ```bash
32
+ pip install git+https://github.com/huggingface/transformers.git
33
+ ```
34
+
35
+ To reproduce our results:
36
+
37
+ ```python
38
+
39
+ from transformers import AutoTokenizer, LlamaForCausalLM
40
+ from human_eval.data import write_jsonl, read_problems
41
+ from tqdm import tqdm
42
+
43
+ # initialize the model
44
+
45
+ model_path = "Phind/Phind-CodeLlama-34B-v1"
46
+ model = LlamaForCausalLM.from_pretrained(model_path, device_map="auto")
47
+ tokenizer = AutoTokenizer.from_pretrained(model_path)
48
+
49
+ # HumanEval helper
50
+
51
+ def generate_one_completion(prompt: str):
52
+ tokenizer.pad_token = tokenizer.eos_token
53
+ inputs = tokenizer(prompt, return_tensors="pt", truncation=True, max_length=4096)
54
+
55
+ # Generate
56
+ generate_ids = model.generate(inputs.input_ids.to("cuda"), max_new_tokens=256, do_sample=True, top_p=0.75, top_k=40, temperature=0.1)
57
+ completion = tokenizer.batch_decode(generate_ids, skip_special_tokens=True, clean_up_tokenization_spaces=False)[0]
58
+ completion = completion.replace(prompt, "").split("\n\n\n")[0]
59
+
60
+ return completion
61
+
62
+ # perform HumanEval
63
+ problems = read_problems()
64
+
65
+ num_samples_per_task = 1
66
+ samples = [
67
+ dict(task_id=task_id, completion=generate_one_completion(problems[task_id]["prompt"]))
68
+ for task_id in tqdm(problems)
69
+ for _ in range(num_samples_per_task)
70
+ ]
71
+ write_jsonl("samples.jsonl", samples)
72
+
73
+ # run `evaluate_functional_correctness samples.jsonl` in your HumanEval code sandbox
74
+ ```
75
+
76
+ ## Bias, Risks, and Limitations
77
+
78
+ <!-- This section is meant to convey both technical and sociotechnical limitations. -->
79
+ This model has undergone very limited testing. Additional safety testing should be performed before any real-world deployments.
80
+
81
+
82
+ ## Training details
83
+
84
+ <!-- Total emissions (in grams of CO2eq) and additional considerations, such as electricity usage, go here. Edit the suggested text below accordingly -->
85
+
86
+ - **Hardware Type:** 32x A100-80GB
87
+ - **Hours used:** 90 GPU-hours
88
+ - **Cloud Provider:** AWS
89
+ - **Compute Region:** us-east-1
90
+