PEFT
PyTorch
gpt_bigcode
code
terryyz commited on
Commit
5be4918
1 Parent(s): 085737f

Upload README.md with huggingface_hub

Browse files
Files changed (1) hide show
  1. README.md +139 -0
README.md CHANGED
@@ -0,0 +1,139 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+
2
+ ---
3
+ license: bigcode-openrail-m
4
+ datasets:
5
+ - bigcode/guanaco-commits
6
+ metrics:
7
+ - code_eval
8
+ library_name: peft
9
+ tags:
10
+ - code
11
+ ---
12
+ # Astraios: A Recipe for Parameter-Efficient Instruction Tuning Code Language Models
13
+ <p align="center" width="100%">
14
+ <a ><img src="https://github.com/bigcode-project/astraios/blob/main/visuals/banner.png?raw=true" alt="Astraios" style="width: 20%; min-width: 300px; display: block; margin: auto;"></a>
15
+ </p>
16
+
17
+ # Table of Contents
18
+
19
+ 1. [Model Summary](#model-summary)
20
+ 2. [Use](#use)
21
+ 3. [Training](#training)
22
+ 4. [Citation](#citation)
23
+
24
+ # Model Summary
25
+
26
+ > Astraios-1B-FFT is an instruction tuned model with 15.5B parameters created by finetuning StarCoderBase on CommitPackFT & OASST as described in the Astraios paper.
27
+
28
+ - **Repository:** [bigcode-project/astraios](https://github.com/bigcode-project/astraios)
29
+ - **Paper:** [Astraios: A Recipe for Parameter Efficient Instruction Tuning Code Language Models]()
30
+ - **Languages:** 80+ Programming languages
31
+ - **✨Astraios:**
32
+ <table>
33
+ <tr>
34
+ <th>Data</t>
35
+ <td><a href=https://huggingface.co/datasets/bigcode/guanaco-commits>CommitPackFT+OASST</a></td>
36
+ <td>Filtered version of CommitPack and OASST for high-quality commit messages that resemble instructions</td>
37
+ </tr>
38
+ <tr>
39
+ <th>Model</t>
40
+ <td><a href=https://huggingface.co/collections/bigcode/astraios-1b-6576ff1b8e449026ae327c1c>Astraios-1B</a></td>
41
+ <td>Collection of StarCoderBase-1B models instruction tuned on CommitPackFT + OASST with different tuning methods</td>
42
+ </tr>
43
+ <tr>
44
+ <th></t>
45
+ <td><a href=https://huggingface.co/collections/bigcode/astraios-3b-6577127317ee44ff547252d3>Astraios-3B</a></td>
46
+ <td>Collection of StarCoderBase-3B (3B parameters) models instruction tuned on CommitPackFT + OASST with different tuning methods</td>
47
+ </tr>
48
+ <tr>
49
+ <th></t>
50
+ <td><a href=https://huggingface.co/collections/starpeft/starcoderbase-7b-650c1f028b45cfec8e72c265>Astraios-7B</a></td>
51
+ <td>Collection of StarCoderBase-7B (7B parameters) models instruction tuned on CommitPackFT + OASST with different tuning methods</td>
52
+ </tr>
53
+ <tr>
54
+ <th></t>
55
+ <td><a href=https://huggingface.co/collections/bigcode/astraios-16b-65788b7476b6de79781054cc>Astraios-16B</a></td>
56
+ <td>Collection of StarCoderBase-16B (16B parameters) models instruction tuned on CommitPackFT + OASST with different tuning methods</td>
57
+ </tr>
58
+ <tr>
59
+ <th>Evaluation</t>
60
+ <td><a href=https://huggingface.co/datasets/code_x_glue_cc_clone_detection_big_clone_bench>BigCloneBench</a></td>
61
+ <td>Dataset for clone detection; We use 2,000 samples for evaluation</td>
62
+ </tr>
63
+ <tr>
64
+ <th></t>
65
+ <td><a href=https://huggingface.co/datasets/code_x_glue_cc_defect_detection>Devign</a></td>
66
+ <td>Dataset for defect detection; We use 2,000 samples for evaluation</td>
67
+ </tr>
68
+ <tr>
69
+ <th></t>
70
+ <td><a href=https://huggingface.co/datasets/bigcode/humanevalpack>HumanEvalPack</a></td>
71
+ <td>Extension of OpenAI's HumanEval to cover 3 scenarios across 6 languages</td>
72
+ </tr>
73
+ <tr>
74
+ <th></t>
75
+ <td><a href=https://huggingface.co/datasets/RaymondLi/perturbed_humaneval>ReCode</a></td>
76
+ <td>Dataset for the robustness of code generation, covering 4 variants</td>
77
+ </tr>
78
+ <tr>
79
+ <th></t>
80
+ <td><a href=https://huggingface.co/datasets/moyix/asleep_keyboard>Asleep At The Keyboard</a></td>
81
+ <td>Datasets for security of code generation; We use DoW for evaluation</td>
82
+ </tr>
83
+ </table>
84
+
85
+
86
+ # Use
87
+
88
+ ## Intended use
89
+
90
+ The model follows instructions provided in the input. You should always preface your input with "Question: " and finish it with "Answer:", for example: "Question: Please write a function in Python that performs bubble sort.
91
+
92
+ Answer:"
93
+
94
+ **Feel free to share your generations in the Community tab!**
95
+
96
+ ## Generation
97
+ ```python
98
+ # pip install -q transformers
99
+ from transformers import AutoModelForCausalLM, AutoTokenizer
100
+
101
+ checkpoint = "bigcode/astraios-1b-fft"
102
+ model = AutoModelForCausalLM.from_pretrained(checkpoint)
103
+ device = "cuda" # for GPU usage or "cpu" for CPU usage
104
+
105
+ tokenizer = AutoTokenizer.from_pretrained(checkpoint)
106
+ model = AutoModelForCausalLM.from_pretrained(checkpoint).to(device)
107
+
108
+ inputs = tokenizer.encode("Question: Please write a function in Python that performs bubble sort.
109
+
110
+ Answer:", return_tensors="pt").to(device)
111
+ outputs = model.generate(inputs)
112
+ print(tokenizer.decode(outputs[0]))
113
+ ```
114
+
115
+ # Training
116
+
117
+ ## Model
118
+
119
+ - **Architecture:** GPT-2 model with multi-query attention and Fill-in-the-Middle objective
120
+ - **Steps:** 250k pretraining & 200 instruction tuning
121
+ - **Precision:** fp32
122
+
123
+ ## Hardware
124
+
125
+ - **Pretraining:**
126
+ - **GPUs:** 512 Tesla A100
127
+ - **Training time:** 24 days
128
+ - **Instruction tuning:**
129
+ - **GPUs:** 8 Tesla A100
130
+
131
+ ## Software
132
+
133
+ - **Orchestration:** [Megatron-LM/Transformers](https://github.com/bigcode-project/octopack#training)
134
+ - **Neural networks:** [PyTorch](https://github.com/pytorch/pytorch)
135
+
136
+ # Citation
137
+
138
+ ```bibtex
139
+ ```