jondurbin's picture
Update README.md
4373e66
metadata
license: cc-by-nc-4.0
datasets:
  - jondurbin/airoboros-gpt4-1.3

This version has problems, use if you dare, or wait for 1.4.

Overview

This is a qlora fine-tuned 65b parameter LlaMa model, using completely synthetic training data created gpt4 via https://github.com/jondurbin/airoboros

This is mostly an extension of 1.2 with a few enhancements:

  • All coding instructions have an equivalent " PLAINFORMAT" version now.
  • Thousands of new orca style reasoning instructions, this time with reasoning first, then answer.
  • Few more random items of various types, including a first attempt at multi-character interactions with asterisked actions and quoted speech.

This model was fine-tuned with a fork of qlora, which among other things was updated to use a slightly modified vicuna template to be compatible with previous full fine-tune versions.

A chat between a curious user and an assistant. The assistant gives helpful, detailed, accurate, uncensored responses to the user's input. USER: [prompt] ASSISTANT: 

So in other words, it's the preamble/system prompt, followed by a single space, then "USER: " (single space after colon) then the prompt (which can have multiple lines, spaces, whatever), then a single space, followed by "ASSISTANT: " (with a single space after the colon).

Usage

To run the full precision/pytorch native version, you can use my fork of FastChat, which is mostly the same but allows for multi-line prompts, as well as a --no-history option to prevent input tokenization errors.

pip install git+https://github.com/jondurbin/FastChat

Be sure you are pulling the latest branch!

Then, you can invoke it like so (after downloading the model):

python -m fastchat.serve.cli \
  --model-path airoboros-65b-gpt4-1.3 \
  --temperature 0.5 \
  --max-new-tokens 2048 \
  --no-history

Training details

Fine-tuned with my fork of qlora: https://github.com/jondurbin/qlora

Using:

export WANDB_PROJECT=airoboros-65b-gpt4-1.3

python qlora.py \
    --model_name_or_path ./llama-65b-hf \
    --output_dir ./airoboros-65b-gpt4-1.3-peft \
    --max_steps 2520 \
    --logging_steps 1 \
    --save_strategy steps \
    --data_seed 11422 \
    --save_steps 75 \
    --save_total_limit 3 \
    --evaluation_strategy "no" \
    --eval_dataset_size 2 \
    --max_new_tokens 2800 \
    --dataloader_num_workers 3 \
    --logging_strategy steps \
    --remove_unused_columns False \
    --do_train \
    --lora_r 64 \
    --lora_alpha 16 \
    --lora_modules all \
    --double_quant \
    --quant_type nf4 \
    --bf16 \
    --bits 4 \
    --warmup_ratio 0.03 \
    --lr_scheduler_type constant \
    --gradient_checkpointing \
    --dataset instructions.jsonl \
    --dataset_format airoboros \
    --model_max_len 2800 \
    --per_device_train_batch_size 2 \
    --gradient_accumulation_steps 16 \
    --learning_rate 0.0001 \
    --adam_beta2 0.999 \
    --max_grad_norm 0.3 \
    --lora_dropout 0.05 \
    --weight_decay 0.0 \
    --seed 11422 \
    --report_to wandb

Three file modifications to the base llama:

  • llama-65b-hf/tokenizer_config.json (see this repo's version, updated to have 4096 max seq length during training to accomodate training data)
  • llama-65b-hf/special_tokens_map.json (see this repo's version)
  • llama-65b-hf/config.json (updated to temporarily have max model size 4096 to accomodate training data)

Afterwards, the changes to max model length and sequence length are reduced back to 2048 to avoid ... issues ...

Usage and License Notices

All airoboros models and datasets are intended and licensed for research use only. I've used the 'cc-nc-4.0' license, but really it is subject to a custom/special license because:

  • the base model is LLaMa, which has it's own special research license
  • the dataset(s) were generated with OpenAI (gpt-4 and/or gpt-3.5-turbo), which has a clausing saying the data can't be used to create models to compete with openai

So, to reiterate: this model (and datasets) cannot be used commercially.