模型:
TheBloke/airoboros-13B-HF
Chat & support: my new Discord server
Want to contribute? TheBloke's Patreon page
这些文件是 Jon Durbin's Airoboros 13B 的HF格式fp16模型文件。
这是将Jon的fp32存储库转换为fp16以便更易于存储和使用的结果。
如需进一步支持以及讨论有关这些模型和人工智能的问题,请加入我们的群组:
感谢 chirper.ai 团队!
有很多人询问是否可以做出贡献。我喜欢提供模型并帮助人们,非常乐意花更多时间进行这样的工作,以及扩展到新的项目,如微调/训练。
如果您有能力和意愿提供帮助,我会非常感激,并且这将帮助我继续提供更多模型,开始新的人工智能项目。
贡献者将优先获得关于任何和所有人工智能/LLM/模型问题和请求的支持,可以进入私人Discord聊天室,以及其他优势。
Patreon特别鸣谢:Aemon Algiz, Dmitriy Samsonov, Nathan LeClaire, Trenton Dambrowitz, Mano Prime, David Flickinger, vamX, Nikolai Manek, senxiiz, Khalefa Al-Ahmad, Illia Dulskyi, Jonathan Leane, Talal Aujan, V. Lukas, Joseph William Delisle, Pyrater, Oscar Rangel, Lone Striker, Luke Pendergrass, Eugene Pentland, Sebastain Graf, Johann-Peter Hartman。
感谢所有慷慨的赞助者和捐赠者!
这是一个使用 https://github.com/jondurbin/airoboros 创建的完全合成训练数据进行微调的13b参数LlaMa模型。
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 |
wb-13b-u是Wizard-Vicuna-13b-Uncensored
airoboros-13b | gpt35 | gpt4-x-alpasta-30b | manticore-13b | vicuna-13b-1.1 | wv-13b-u | link |
---|---|---|---|---|---|---|
80 | 95 | 70 | 90 | 85 | 60 | 12314321 |
20 | 95 | 40 | 30 | 90 | 80 | 12315321 |
100 | 100 | 100 | 95 | 95 | 100 | 12316321 |
90 | 100 | 85 | 60 | 95 | 100 | 12317321 |
95 | 90 | 80 | 85 | 95 | 75 | 12318321 |
100 | 95 | 90 | 95 | 98 | 92 | 12319321 |
50 | 100 | 80 | 95 | 60 | 55 | 12320321 |
70 | 90 | 80 | 60 | 85 | 40 | 12321321 |
100 | 95 | 50 | 85 | 40 | 60 | 12322321 |
85 | 60 | 55 | 65 | 50 | 70 | 12323321 |
95 | 100 | 85 | 90 | 60 | 75 | 12324321 |
100 | 95 | 70 | 80 | 50 | 85 | 12325321 |
100 | 95 | 80 | 70 | 60 | 90 | 12326321 |
95 | 100 | 70 | 85 | 90 | 90 | 12327321 |
80 | 95 | 90 | 60 | 30 | 85 | 12328321 |
60 | 95 | 0 | 75 | 50 | 40 | 12329321 |
100 | 95 | 90 | 98 | 95 | 95 | 12330321 |
60 | 85 | 40 | 50 | 20 | 0 | 12331321 |
100 | 90 | 85 | 95 | 95 | 80 | 12332321 |
100 | 95 | 100 | 95 | 90 | 95 | 12333321 |
95 | 90 | 96 | 80 | 92 | 88 | 12334321 |
95 | 92 | 90 | 93 | 89 | 91 | 12335321 |
95 | 93 | 90 | 94 | 96 | 92 | 12336321 |
95 | 90 | 93 | 88 | 92 | 85 | 12337321 |
95 | 90 | 85 | 96 | 88 | 92 | 12338321 |
95 | 95 | 90 | 93 | 92 | 91 | 12339321 |
95 | 98 | 80 | 97 | 99 | 96 | 12340321 |
95 | 93 | 90 | 87 | 92 | 89 | 12341321 |
90 | 85 | 95 | 80 | 92 | 75 | 12342321 |
90 | 85 | 95 | 93 | 80 | 92 | 12343321 |
95 | 92 | 90 | 91 | 93 | 89 | 12344321 |
100 | 95 | 90 | 85 | 80 | 95 | 12345321 |
95 | 97 | 93 | 92 | 96 | 94 | 12346321 |
95 | 93 | 94 | 90 | 88 | 92 | 12347321 |
90 | 95 | 98 | 85 | 96 | 92 | 12348321 |
90 | 88 | 85 | 80 | 82 | 84 | 12349321 |
90 | 95 | 85 | 87 | 92 | 88 | 12350321 |
95 | 97 | 96 | 90 | 93 | 92 | 12351321 |
95 | 93 | 92 | 90 | 89 | 91 | 12352321 |
90 | 95 | 93 | 92 | 94 | 91 | 12353321 |
90 | 85 | 95 | 80 | 88 | 75 | 12354321 |
85 | 90 | 95 | 88 | 92 | 80 | 12355321 |
90 | 95 | 92 | 85 | 80 | 87 | 12356321 |
85 | 90 | 95 | 80 | 88 | 75 | 12357321 |
85 | 80 | 75 | 90 | 70 | 82 | 12358321 |
90 | 85 | 95 | 92 | 93 | 80 | 12359321 |
90 | 95 | 75 | 85 | 80 | 70 | 12360321 |
85 | 90 | 80 | 88 | 82 | 83 | 12361321 |
85 | 90 | 95 | 92 | 88 | 80 | 12362321 |
85 | 90 | 80 | 75 | 95 | 88 | 12363321 |
85 | 90 | 80 | 88 | 84 | 92 | 12364321 |
80 | 90 | 75 | 85 | 70 | 95 | 12365321 |
90 | 88 | 85 | 80 | 92 | 83 | 12366321 |
85 | 75 | 90 | 80 | 78 | 88 | 12367321 |
85 | 90 | 80 | 82 | 75 | 88 | 12368321 |
90 | 85 | 40 | 95 | 80 | 88 | 12369321 |
85 | 95 | 90 | 75 | 88 | 80 | 12370321 |
85 | 95 | 90 | 92 | 89 | 88 | 12371321 |
80 | 85 | 75 | 60 | 90 | 70 | 12372321 |
85 | 90 | 87 | 80 | 88 | 75 | 12373321 |
85 | 80 | 75 | 50 | 90 | 80 | 12374321 |
95 | 80 | 90 | 85 | 75 | 82 | 12375321 |
85 | 90 | 80 | 70 | 95 | 88 | 12376321 |
90 | 95 | 70 | 85 | 80 | 75 | 12377321 |
90 | 85 | 70 | 75 | 80 | 60 | 12378321 |
95 | 90 | 70 | 50 | 85 | 80 | 12379321 |
80 | 85 | 40 | 60 | 90 | 95 | 12380321 |
75 | 60 | 80 | 55 | 70 | 85 | 12381321 |
90 | 85 | 60 | 50 | 80 | 95 | 12382321 |
45 | 85 | 60 | 20 | 65 | 75 | 12383321 |
85 | 90 | 30 | 60 | 80 | 70 | 12384321 |
90 | 95 | 80 | 40 | 85 | 70 | 12385321 |
85 | 90 | 70 | 75 | 80 | 95 | 12386321 |
90 | 70 | 50 | 20 | 60 | 40 | 12387321 |
90 | 95 | 75 | 60 | 85 | 80 | 12388321 |
85 | 80 | 60 | 70 | 65 | 75 | 12389321 |
90 | 85 | 80 | 75 | 82 | 70 | 12390321 |
90 | 95 | 80 | 70 | 85 | 75 | 12391321 |
85 | 75 | 30 | 80 | 90 | 70 | 12392321 |
85 | 90 | 50 | 70 | 80 | 60 | 12393321 |
100 | 95 | 98 | 99 | 97 | 96 | 12394321 |
95 | 90 | 92 | 93 | 91 | 89 | 12395321 |
95 | 92 | 90 | 85 | 88 | 91 | 12396321 |
100 | 95 | 98 | 97 | 96 | 99 | 12397321 |
100 | 100 | 100 | 90 | 100 | 95 | 12398321 |
100 | 95 | 98 | 97 | 94 | 99 | 12399321 |
95 | 90 | 92 | 93 | 94 | 91 | 123100321 |
100 | 95 | 98 | 90 | 96 | 95 | 123101321 |
95 | 96 | 92 | 90 | 89 | 93 | 123102321 |
100 | 95 | 93 | 90 | 92 | 88 | 123103321 |
100 | 100 | 98 | 97 | 99 | 100 | 123104321 |
95 | 90 | 92 | 85 | 93 | 94 | 123105321 |
95 | 93 | 90 | 92 | 96 | 91 | 123106321 |
95 | 96 | 92 | 90 | 93 | 91 | 123107321 |
95 | 90 | 92 | 93 | 91 | 89 | 123108321 |
100 | 98 | 95 | 97 | 96 | 99 | 123109321 |
90 | 95 | 85 | 88 | 92 | 87 | 123110321 |
95 | 93 | 90 | 92 | 89 | 88 | 123111321 |
100 | 95 | 97 | 90 | 96 | 94 | 123112321 |
95 | 93 | 90 | 92 | 94 | 91 | 123113321 |
95 | 92 | 90 | 93 | 94 | 88 | 123114321 |
95 | 92 | 60 | 97 | 90 | 96 | 123115321 |
95 | 90 | 92 | 93 | 91 | 89 | 123116321 |
95 | 90 | 97 | 92 | 91 | 93 | 123117321 |
90 | 95 | 93 | 85 | 92 | 91 | 123118321 |
95 | 90 | 40 | 92 | 93 | 85 | 123119321 |
100 | 100 | 95 | 90 | 95 | 90 | 123120321 |
90 | 95 | 96 | 98 | 93 | 92 | 123121321 |
90 | 95 | 92 | 89 | 93 | 94 | 123122321 |
100 | 95 | 100 | 98 | 96 | 99 | 123123321 |
100 | 100 | 95 | 90 | 100 | 90 | 123124321 |
90 | 85 | 88 | 92 | 87 | 91 | 123125321 |
95 | 97 | 90 | 92 | 93 | 94 | 123126321 |
90 | 95 | 85 | 88 | 92 | 89 | 123127321 |
95 | 93 | 90 | 92 | 94 | 91 | 123128321 |
90 | 95 | 85 | 80 | 88 | 82 | 123129321 |
95 | 90 | 60 | 85 | 93 | 70 | 123130321 |
95 | 92 | 94 | 93 | 96 | 90 | 123131321 |
95 | 90 | 85 | 93 | 87 | 92 | 123132321 |
95 | 96 | 93 | 90 | 97 | 92 | 123133321 |
100 | 0 | 0 | 100 | 0 | 0 | 123134321 |
60 | 100 | 0 | 80 | 0 | 0 | 123135321 |
0 | 100 | 60 | 0 | 0 | 90 | 123136321 |
100 | 100 | 0 | 100 | 100 | 100 | 123137321 |
100 | 100 | 100 | 100 | 95 | 100 | 123138321 |
100 | 100 | 100 | 50 | 90 | 100 | 123139321 |
100 | 100 | 100 | 100 | 95 | 90 | 123140321 |
100 | 100 | 100 | 95 | 0 | 100 | 123141321 |
50 | 95 | 20 | 10 | 30 | 85 | 123142321 |
100 | 100 | 60 | 20 | 30 | 40 | 123143321 |
100 | 0 | 0 | 0 | 0 | 100 | 123144321 |
0 | 100 | 60 | 0 | 0 | 80 | 123145321 |
50 | 100 | 20 | 90 | 0 | 10 | 123146321 |
100 | 100 | 100 | 100 | 100 | 100 | 123147321 |
100 | 100 | 100 | 100 | 100 | 100 | 123148321 |
40 | 100 | 95 | 0 | 100 | 40 | 123149321 |
100 | 100 | 100 | 100 | 80 | 100 | 123150321 |
100 | 100 | 100 | 0 | 90 | 40 | 123151321 |
0 | 100 | 100 | 50 | 70 | 20 | 123152321 |
100 | 100 | 50 | 90 | 0 | 95 | 123153321 |
100 | 95 | 90 | 85 | 98 | 80 | 123154321 |
95 | 98 | 90 | 92 | 96 | 89 | 123155321 |
90 | 95 | 75 | 85 | 80 | 82 | 123156321 |
95 | 98 | 50 | 92 | 96 | 94 | 123157321 |
95 | 90 | 0 | 93 | 92 | 94 | 123158321 |
95 | 90 | 85 | 92 | 80 | 88 | 123159321 |
95 | 93 | 75 | 85 | 90 | 92 | 123160321 |
90 | 95 | 88 | 85 | 92 | 89 | 123161321 |
100 | 100 | 100 | 95 | 97 | 98 | 123162321 |
85 | 40 | 30 | 95 | 90 | 88 | 123163321 |
90 | 95 | 92 | 85 | 88 | 93 | 123164321 |
95 | 96 | 92 | 90 | 89 | 93 | 123165321 |
90 | 95 | 85 | 80 | 92 | 88 | 123166321 |
95 | 98 | 65 | 90 | 85 | 93 | 123167321 |
95 | 92 | 96 | 97 | 90 | 89 | 123168321 |
95 | 90 | 92 | 91 | 89 | 93 | 123169321 |
95 | 90 | 80 | 75 | 95 | 90 | 123170321 |
92 | 40 | 30 | 95 | 90 | 93 | 123171321 |
90 | 92 | 85 | 88 | 89 | 87 | 123172321 |
95 | 80 | 90 | 92 | 91 | 88 | 123173321 |
95 | 93 | 92 | 90 | 91 | 94 | 123174321 |
100 | 98 | 95 | 90 | 92 | 96 | 123175321 |
95 | 92 | 80 | 85 | 90 | 93 | 123176321 |
95 | 98 | 90 | 88 | 97 | 96 | 123177321 |
90 | 95 | 85 | 88 | 86 | 92 | 123178321 |
100 | 100 | 100 | 100 | 100 | 100 | 123179321 |
90 | 95 | 85 | 96 | 92 | 88 | 123180321 |
100 | 98 | 95 | 99 | 97 | 96 | 123181321 |
95 | 92 | 70 | 90 | 93 | 89 | 123182321 |
95 | 90 | 88 | 92 | 94 | 93 | 123183321 |
95 | 90 | 93 | 92 | 85 | 94 | 123184321 |
95 | 93 | 90 | 87 | 92 | 91 | 123185321 |
95 | 93 | 90 | 96 | 92 | 91 | 123186321 |
95 | 97 | 85 | 96 | 98 | 90 | 123187321 |
95 | 92 | 90 | 85 | 93 | 94 | 123188321 |
95 | 96 | 92 | 90 | 97 | 93 | 123189321 |
95 | 93 | 96 | 94 | 90 | 92 | 123190321 |
95 | 94 | 93 | 92 | 90 | 89 | 123191321 |
90 | 85 | 95 | 80 | 87 | 75 | 123192321 |
95 | 94 | 92 | 93 | 90 | 96 | 123193321 |
95 | 100 | 90 | 95 | 95 | 95 | 123194321 |
100 | 95 | 85 | 100 | 0 | 90 | 123195321 |
100 | 95 | 90 | 95 | 100 | 95 | 123196321 |
95 | 90 | 60 | 95 | 85 | 80 | 123197321 |
100 | 95 | 90 | 98 | 97 | 99 | 123198321 |
95 | 90 | 85 | 95 | 80 | 92 | 123199321 |
100 | 95 | 100 | 98 | 100 | 90 | 123200321 |
100 | 95 | 80 | 85 | 90 | 85 | 123201321 |
100 | 90 | 95 | 85 | 95 | 100 | 123202321 |
95 | 90 | 85 | 80 | 88 | 92 | 123203321 |
100 | 100 | 0 | 0 | 100 | 0 | 123204321 |
100 | 100 | 100 | 50 | 100 | 75 | 123205321 |
100 | 100 | 0 | 0 | 100 | 0 | 123206321 |
0 | 100 | 0 | 0 | 0 | 0 | 123207321 |
100 | 100 | 50 | 0 | 0 | 0 | 123208321 |
100 | 100 | 100 | 100 | 100 | 95 | 123209321 |
100 | 100 | 50 | 0 | 0 | 0 | 123210321 |
100 | 100 | 0 | 0 | 100 | 0 | 123211321 |
90 | 85 | 80 | 95 | 70 | 75 | 123212321 |
100 | 100 | 0 | 0 | 0 | 0 | 123213321 |
我使用一个越狱提示生成合成指令,这导致一些训练数据可能会被其他模型审查,例如有关合成药物、制作家庭自制喷火器等的操作提示。请注意,这些都是由ChatGPT生成的,而不是我个人创建的。我的目标只是为了测试ChatGPT在未经过滤(尽可能多)的情况下的一些能力,而不是故意产生任何有害/危险等内容。
我使用的越狱提示是在使用 --uncensored 标志时python代码的默认提示: https://github.com/jondurbin/airoboros/blob/main/airoboros/self_instruct.py#L39
我还进行了几次手动清理,以删除一些不良提示,但基本上我将数据保持不变。最初,该模型在数学/外推、封闭问题回答(大量幻觉)和编码方面表现得相当糟糕,因此我进行了一次精细调整,使用了其他合成指令来解决这些类型的问题。
初始指令和最终精细调整指令将很快发布。
我使用了优秀的 FastChat 模块进行微调,运行命令如下:
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
这在8个NVIDIA 80GB A100上运行了约40小时。
提示应与FastChat/vicuna格式完全兼容,例如:
带前导语:
A chat between a curious user and an artificial intelligence assistant. The assistant gives helpful, detailed, and polite answers to the user's questions. USER: [prompt] <\s> ASSISTANT:
或者只是:
USER: [prompt] <\s> ASSISTANT:
该模型使用LLaMA模型许可,数据集使用OpenAI条款许可,其他所有内容均可免费使用。