airoboros-13b / README.md
jondurbin's picture
Update README.md
17d39e2
|
raw
history blame
38.6 kB
metadata
license: other

Overview

This is a fine-tuned 13b parameter LlaMa model, using completely synthetic training data created by https://github.com/jondurbin/airoboros

Eval (gpt4 judging)

model raw score gpt-3.5 adjusted score
airoboros-13b 17947 98.087
gpt35 18297 100.0
gpt4-x-alpasta-30b 15612 85.33
manticore-13b 15856 86.66
vicuna-13b-1.1 16306 89.12
wizard-vicuna-13b-uncensored 16287 89.01
question airoboros-13b gpt35 gpt4-x-alpasta-30b manticore-13b vicuna-13b-1.1 wizard-vicuna-13b-uncensored link
1 80 95 70 90 85 60 eval
2 20 95 40 30 90 80 eval
3 100 100 100 95 95 100 eval
4 90 100 85 60 95 100 eval
5 95 90 80 85 95 75 eval
6 100 95 90 95 98 92 eval
7 50 100 80 95 60 55 eval
8 70 90 80 60 85 40 eval
9 100 95 50 85 40 60 eval
10 85 60 55 65 50 70 eval
11 95 100 85 90 60 75 eval
12 100 95 70 80 50 85 eval
13 100 95 80 70 60 90 eval
14 95 100 70 85 90 90 eval
15 80 95 90 60 30 85 eval
16 60 95 0 75 50 40 eval
17 100 95 90 98 95 95 eval
18 60 85 40 50 20 0 eval
19 100 90 85 95 95 80 eval
20 100 95 100 95 90 95 eval
21 95 90 96 80 92 88 eval
22 95 92 90 93 89 91 eval
23 95 93 90 94 96 92 eval
24 95 90 93 88 92 85 eval
25 95 90 85 96 88 92 eval
26 95 95 90 93 92 91 eval
27 95 98 80 97 99 96 eval
28 95 93 90 87 92 89 eval
29 90 85 95 80 92 75 eval
30 90 85 95 93 80 92 eval
31 95 92 90 91 93 89 eval
32 100 95 90 85 80 95 eval
33 95 97 93 92 96 94 eval
34 95 93 94 90 88 92 eval
35 90 95 98 85 96 92 eval
36 90 88 85 80 82 84 eval
37 90 95 85 87 92 88 eval
38 95 97 96 90 93 92 eval
39 95 93 92 90 89 91 eval
40 90 95 93 92 94 91 eval
41 90 85 95 80 88 75 eval
42 85 90 95 88 92 80 eval
43 90 95 92 85 80 87 eval
44 85 90 95 80 88 75 eval
45 85 80 75 90 70 82 eval
46 90 85 95 92 93 80 eval
47 90 95 75 85 80 70 eval
48 85 90 80 88 82 83 eval
49 85 90 95 92 88 80 eval
50 85 90 80 75 95 88 eval
51 85 90 80 88 84 92 eval
52 80 90 75 85 70 95 eval
53 90 88 85 80 92 83 eval
54 85 75 90 80 78 88 eval
55 85 90 80 82 75 88 eval
56 90 85 40 95 80 88 eval
57 85 95 90 75 88 80 eval
58 85 95 90 92 89 88 eval
59 80 85 75 60 90 70 eval
60 85 90 87 80 88 75 eval
61 85 80 75 50 90 80 eval
62 95 80 90 85 75 82 eval
63 85 90 80 70 95 88 eval
64 90 95 70 85 80 75 eval
65 90 85 70 75 80 60 eval
66 95 90 70 50 85 80 eval
67 80 85 40 60 90 95 eval
68 75 60 80 55 70 85 eval
69 90 85 60 50 80 95 eval
70 45 85 60 20 65 75 eval
71 85 90 30 60 80 70 eval
72 90 95 80 40 85 70 eval
73 85 90 70 75 80 95 eval
74 90 70 50 20 60 40 eval
75 90 95 75 60 85 80 eval
76 85 80 60 70 65 75 eval
77 90 85 80 75 82 70 eval
78 90 95 80 70 85 75 eval
79 85 75 30 80 90 70 eval
80 85 90 50 70 80 60 eval
81 100 95 98 99 97 96 eval
82 95 90 92 93 91 89 eval
83 95 92 90 85 88 91 eval
84 100 95 98 97 96 99 eval
85 100 100 100 90 100 95 eval
86 100 95 98 97 94 99 eval
87 95 90 92 93 94 91 eval
88 100 95 98 90 96 95 eval
89 95 96 92 90 89 93 eval
90 100 95 93 90 92 88 eval
91 100 100 98 97 99 100 eval
92 95 90 92 85 93 94 eval
93 95 93 90 92 96 91 eval
94 95 96 92 90 93 91 eval
95 95 90 92 93 91 89 eval
96 100 98 95 97 96 99 eval
97 90 95 85 88 92 87 eval
98 95 93 90 92 89 88 eval
99 100 95 97 90 96 94 eval
100 95 93 90 92 94 91 eval
101 95 92 90 93 94 88 eval
102 95 92 60 97 90 96 eval
103 95 90 92 93 91 89 eval
104 95 90 97 92 91 93 eval
105 90 95 93 85 92 91 eval
106 95 90 40 92 93 85 eval
107 100 100 95 90 95 90 eval
108 90 95 96 98 93 92 eval
109 90 95 92 89 93 94 eval
110 100 95 100 98 96 99 eval
111 100 100 95 90 100 90 eval
112 90 85 88 92 87 91 eval
113 95 97 90 92 93 94 eval
114 90 95 85 88 92 89 eval
115 95 93 90 92 94 91 eval
116 90 95 85 80 88 82 eval
117 95 90 60 85 93 70 eval
118 95 92 94 93 96 90 eval
119 95 90 85 93 87 92 eval
120 95 96 93 90 97 92 eval
121 100 0 0 100 0 0 eval
122 60 100 0 80 0 0 eval
123 0 100 60 0 0 90 eval
124 100 100 0 100 100 100 eval
125 100 100 100 100 95 100 eval
126 100 100 100 50 90 100 eval
127 100 100 100 100 95 90 eval
128 100 100 100 95 0 100 eval
129 50 95 20 10 30 85 eval
130 100 100 60 20 30 40 eval
131 100 0 0 0 0 100 eval
132 0 100 60 0 0 80 eval
133 50 100 20 90 0 10 eval
134 100 100 100 100 100 100 eval
135 100 100 100 100 100 100 eval
136 40 100 95 0 100 40 eval
137 100 100 100 100 80 100 eval
138 100 100 100 0 90 40 eval
139 0 100 100 50 70 20 eval
140 100 100 50 90 0 95 eval
141 100 95 90 85 98 80 eval
142 95 98 90 92 96 89 eval
143 90 95 75 85 80 82 eval
144 95 98 50 92 96 94 eval
145 95 90 0 93 92 94 eval
146 95 90 85 92 80 88 eval
147 95 93 75 85 90 92 eval
148 90 95 88 85 92 89 eval
149 100 100 100 95 97 98 eval
150 85 40 30 95 90 88 eval
151 90 95 92 85 88 93 eval
152 95 96 92 90 89 93 eval
153 90 95 85 80 92 88 eval
154 95 98 65 90 85 93 eval
155 95 92 96 97 90 89 eval
156 95 90 92 91 89 93 eval
157 95 90 80 75 95 90 eval
158 92 40 30 95 90 93 eval
159 90 92 85 88 89 87 eval
160 95 80 90 92 91 88 eval
161 95 93 92 90 91 94 eval
162 100 98 95 90 92 96 eval
163 95 92 80 85 90 93 eval
164 95 98 90 88 97 96 eval
165 90 95 85 88 86 92 eval
166 100 100 100 100 100 100 eval
167 90 95 85 96 92 88 eval
168 100 98 95 99 97 96 eval
169 95 92 70 90 93 89 eval
170 95 90 88 92 94 93 eval
171 95 90 93 92 85 94 eval
172 95 93 90 87 92 91 eval
173 95 93 90 96 92 91 eval
174 95 97 85 96 98 90 eval
175 95 92 90 85 93 94 eval
176 95 96 92 90 97 93 eval
177 95 93 96 94 90 92 eval
178 95 94 93 92 90 89 eval
179 90 85 95 80 87 75 eval
180 95 94 92 93 90 96 eval
181 95 100 90 95 95 95 eval
182 100 95 85 100 0 90 eval
183 100 95 90 95 100 95 eval
184 95 90 60 95 85 80 eval
185 100 95 90 98 97 99 eval
186 95 90 85 95 80 92 eval
187 100 95 100 98 100 90 eval
188 100 95 80 85 90 85 eval
189 100 90 95 85 95 100 eval
190 95 90 85 80 88 92 eval
191 100 100 0 0 100 0 eval
192 100 100 100 50 100 75 eval
193 100 100 0 0 100 0 eval
194 0 100 0 0 0 0 eval
195 100 100 50 0 0 0 eval
196 100 100 100 100 100 95 eval
197 100 100 50 0 0 0 eval
198 100 100 0 0 100 0 eval
199 90 85 80 95 70 75 eval
200 100 100 0 0 0 0 eval

Training data

I used a jailbreak prompt to generate the synthetic instructions, which resulted in some training data that would likely be censored by other models, such as how-to prompts about synthesizing drugs, making homemade flamethrowers, etc. Mind you, this is all generated by ChatGPT, not me. My goal was to simply test some of the capabilities of ChatGPT when unfiltered (as much as possible), and not to intentionally produce any harmful/dangerous/etc. content.

The jailbreak prompt I used is the default prompt in the python code when using the --uncensored flag: https://github.com/jondurbin/airoboros/blob/main/airoboros/self_instruct.py#L39

I also did a few passes of manually cleanup to remove some bad prompts, but mostly I left the data as-is. Initially, the model was fairly bad at math/extrapolation, closed question-answering (heavy hallucination), and coding, so I did one more fine tuning pass with additional synthetic instructions aimed at those types of problems.

Both the initial instructions and final-pass fine-tuning instructions will be published soon.

Fine-tuning method

I used the excellent FastChat module, running with:

source /workspace/venv/bin/activate

export NCCL_P2P_DISABLE=1
export NCCL_P2P_LEVEL=LOC

torchrun --nproc_per_node=8 --master_port=20001 /workspace/FastChat/fastchat/train/train_mem.py \
  --model_name_or_path /workspace/llama-13b \
  --data_path /workspace/as_conversations.json \
  --bf16 True \
  --output_dir /workspace/airoboros-uncensored-13b \
  --num_train_epochs 3 \
  --per_device_train_batch_size 20 \
  --per_device_eval_batch_size 20 \
  --gradient_accumulation_steps 2 \
  --evaluation_strategy "steps" \
  --eval_steps 500 \
  --save_strategy "steps" \
  --save_steps 500 \
  --save_total_limit 10 \
  --learning_rate 2e-5 \
  --weight_decay 0. \
  --warmup_ratio 0.04 \
  --lr_scheduler_type "cosine" \
  --logging_steps 1 \
  --fsdp "full_shard auto_wrap offload" \
  --fsdp_transformer_layer_cls_to_wrap 'LlamaDecoderLayer' \
  --tf32 True \
  --model_max_length 2048 \
  --gradient_checkpointing True \
  --lazy_preprocess True

This ran on 8x nvidia 80gb a100's for about 40 hours.

License

The model is licensed under the LLaMA model, and the dataset is licensed under the terms of OpenAI because it uses ChatGPT. Everything else is free.