google

/

gemma-7b-it

like 1.08k

I'm not too familiar with this code, is this in your custom codebase?

cc @lewtun @lvwerra

Hello @Mansikalra it seems the dataset you're evaluating on doesn't have a test split which is assumed to exist in the SFT example in TRL. You can either comment out the eval_dataset lines of the script or use a dataset with a test split like this for example:

accelerate launch --config_file=examples/accelerate_configs/multi_gpu.yaml examples/scripts/sft.py \
    --model_name google/gemma-7b \
    --dataset_name OpenAssistant/oasst_top1_2023-08-25 \
    --output_dir scratch/gemma-finetuned-oasst \
    --load_in_4bit --use_peft \
    --per_device_train_batch_size 4 \
    --lora_target_modules gate_proj up_proj down_proj q_proj k_proj v_proj o_proj

Thanks a lot, @lewtun , code sample working perfectly.
Initially, was using stingning/ultrachat -->
https://huggingface.co/datasets/stingning/ultrachat/tree/main

But it wasn't working with code :

python trl/examples/scripts/sft.py
--model_name google/gemma-7b
--dataset_name stingning/ultrachat
--load_in_4bit
--use_peft
--batch_size 4
--gradient_accumulation_steps 2 \

Ah yes indeed that dataset isn't pre-formatted for TRL! Here's a smaller subset that also should work: https://huggingface.co/datasets/trl-lib/ultrachat_200k_chatml