Slightly modified mpt-30b, which has some updates to allow gradient checkpointing/etc., to be compatible with qlora training code.
Original model: https://huggingface.co/mosaicml/mpt-30b
My fork of qlora with mpt-30b support: https://github.com/jondurbin/qlora
Differences in the qlora scripts:
- requires adding
--mpt Truefor mpt-based models - uses
--num_train_epochsinstead of--max_steps - uses airoboros prompt format (mostly 1:1 with vicuna) rather than alpaca, and expects an input file in JSONL format with "instruction" and "response"
I think there's a bug in gradient accumulation, so if you try this, maybe set gradient accumulation steps to 1
my first attempts used batch size 6, with gradient accumulation steps 16, but results of three epochs with gradient accumulation vs without were quite a bit worse
5 epochs seemed to achieve the best results, but YMMV
Full example of tuning (used for airoboros-mpt-30b-gpt4-1.4):
source /workspace/venv/bin/activate
export PYTHONPATH=./mpt-30b
export WANDB_API_KEY=[redacted]
export WANDB_PROJECT=airoboros-mpt-30b-gpt4-1.4
python qlora.py \
--model_name_or_path ./mpt-30b \
--output_dir ./$WANDB_PROJECT-checkpoints \
--num_train_epochs 5 \
--logging_steps 1 \
--save_strategy steps \
--data_seed 11422 \
--save_steps 100 \
--save_total_limit 3 \
--evaluation_strategy "no" \
--eval_dataset_size 2 \
--max_new_tokens 8192 \
--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 \
--dataset ./instructions.jsonl \
--dataset_format airoboros \
--model_max_len 8192 \
--gradient_checkpointing \
--per_device_train_batch_size 6 \
--gradient_accumulation_steps 1 \
--learning_rate 0.0001 \
--adam_beta2 0.999 \
--max_grad_norm 0.3 \
--lora_dropout 0.05 \
--weight_decay 0.0 \
--seed 11422 \
--trust_remote_code \
--mpt True \
--report_to wandb
Merged model
Run the merge_weights.py script in the qlora repo: https://github.com/jondurbin/qlora/blob/main/merge_weights.py
Then, copy all of the original python files from the mpt-30b repo into your output directory: https://huggingface.co/mosaicml/mpt-30b/tree/main