StableVicuna-13B is a
Vicuna-13B v0 model fine-tuned using reinforcement learning from human feedback (RLHF) via Proximal Policy Optimization (PPO) on various conversational and instructional datasets.
StableVicuna-13B cannot be used from the
CarperAI/stable-vicuna-13b-delta weights alone. To obtain the correct model, one must add back the difference between LLaMA 13B and
CarperAI/stable-vicuna-13b-delta weights. We provide the
apply_delta.py script to automate the conversion, which you can run as:
Once the delta weights are applied, get started chatting with the model by using the
transformers library. Following a suggestion from Vicuna Team with Vicuna v0 you should install transformers with this version:
1from transformers import AutoTokenizer, AutoModelForCausalLM
2
3tokenizer = AutoTokenizer.from_pretrained("path/to/stable-vicuna-13b-applied")
4model = AutoModelForCausalLM.from_pretrained("path/to/stable-vicuna-13b-applied")
5model.half().cuda()
6
7prompt = """\
8### Human: Write a Python script for text classification using Transformers and PyTorch
9### Assistant:\
10"""
11
12inputs = tokenizer(prompt, return_tensors='pt').to('cuda')
13tokens = model.generate(
14 **inputs,
15 max_new_tokens=256,
16 do_sample=True,
17 temperature=1.0,
18 top_p=1.0,
19)
20print(tokenizer.decode(tokens[0], skip_special_tokens=True))
StableVicuna-13B is fine-tuned on a mix of three datasets.
OpenAssistant Conversations Dataset (OASST1), a human-generated, human-annotated assistant-style conversation corpus consisting of 161,443 messages distributed across 66,497 conversation trees, in 35 different languages;
GPT4All Prompt Generations, a dataset of 400k prompts and responses generated by GPT-4; and
Alpaca, a dataset of 52,000 instructions and demonstrations generated by OpenAI's text-davinci-003 engine.
The reward model used during RLHF was also trained on
OpenAssistant Conversations Dataset (OASST1) along with two other datasets:
Anthropic HH-RLHF, a dataset of preferences about AI assistant helpfulness and harmlessness; and
Stanford Human Preferences Dataset a dataset of 385K collective human preferences over responses to questions/instructions in 18 different subject areas, from cooking to legal advice.
CarperAI/stable-vicuna-13b-delta was trained using PPO as implemented in
trlX with the following configuration:
This model is intended to be used for text generation with a focus on conversational tasks. Users may further fine-tune the model on their own data to improve the model's performance on their specific tasks in accordance with the non-commercial
license.
The base LLaMA model is trained on various data, some of which may contain offensive, harmful, and biased content that can lead to toxic behavior. See Section 5.1 of the LLaMA
paper. We have not performed any studies to determine how fine-tuning on the aforementioned datasets affect the model's behavior and toxicity. Do not treat chat responses from this model as a substitute for human judgment or as a source of truth. Please use responsibly.
This work would not have been possible without the support of
Stability AI.
1@article{touvron2023llama,
2 title={LLaMA: Open and Efficient Foundation Language Models},
3 author={Touvron, Hugo and Lavril, Thibaut and Izacard, Gautier and Martinet, Xavier and Lachaux, Marie-Anne and Lacroix, Timoth{\'e}e and Rozi{\`e}re, Baptiste and Goyal, Naman and Hambro, Eric and Azhar, Faisal and Rodriguez, Aurelien and Joulin, Armand and Grave, Edouard and Lample, Guillaume},
4 journal={arXiv preprint arXiv:2302.13971},
5 year={2023}
6}
1@misc{vicuna2023,
2 title = {Vicuna: An Open-Source Chatbot Impressing GPT-4 with 90%* ChatGPT Quality},
3 url = {https://vicuna.lmsys.org},
4 author = {Chiang, Wei-Lin and Li, Zhuohan and Lin, Zi and Sheng, Ying and Wu, Zhanghao and Zhang, Hao and Zheng, Lianmin and Zhuang, Siyuan and Zhuang, Yonghao and Gonzalez, Joseph E. and Stoica, Ion and Xing, Eric P.},
5 month = {March},
6 year = {2023}
7}
1@misc{gpt4all,
2 author = {Yuvanesh Anand and Zach Nussbaum and Brandon Duderstadt and Benjamin Schmidt and Andriy Mulyar},
3 title = {GPT4All: Training an Assistant-style Chatbot with Large Scale Data Distillation from GPT-3.5-Turbo},
4 year = {2023},
5 publisher = {GitHub},
6 journal = {GitHub repository},
7 howpublished = {\url{https://github.com/nomic-ai/gpt4all}},
8}
1@misc{alpaca,
2 author = {Rohan Taori and Ishaan Gulrajani and Tianyi Zhang and Yann Dubois and Xuechen Li and Carlos Guestrin and Percy Liang and Tatsunori B. Hashimoto },
3 title = {Stanford Alpaca: An Instruction-following LLaMA model},
4 year = {2023},
5 publisher = {GitHub},
6 journal = {GitHub repository},
7 howpublished = {\url{https://github.com/tatsu-lab/stanford_alpaca}},
8}
1@software{leandro_von_werra_2023_7790115,
2 author = {Leandro von Werra and
3 Alex Havrilla and
4 Max reciprocated and
5 Jonathan Tow and
6 Aman cat-state and
7 Duy V. Phung and
8 Louis Castricato and
9 Shahbuland Matiana and
10 Alan and
11 Ayush Thakur and
12 Alexey Bukhtiyarov and
13 aaronrmm and
14 Fabrizio Milo and
15 Daniel and
16 Daniel King and
17 Dong Shin and
18 Ethan Kim and
19 Justin Wei and
20 Manuel Romero and
21 Nicky Pochinkov and
22 Omar Sanseviero and
23 Reshinth Adithyan and
24 Sherman Siu and
25 Thomas Simonini and
26 Vladimir Blagojevic and
27 Xu Song and
28 Zack Witten and
29 alexandremuzio and
30 crumb},
31 title = {{CarperAI/trlx: v0.6.0: LLaMa (Alpaca), Benchmark
32 Util, T5 ILQL, Tests}},
33 month = mar,
34 year = 2023,
35 publisher = {Zenodo},
36 version = {v0.6.0},
37 doi = {10.5281/zenodo.7790115},
38 url = {https://doi.org/10.5281/zenodo.7790115}
39}
