dball/zephyr-7b-sft-qlora

zephyr-7b-sft-qlora

This model is a fine-tuned version of mistralai/Mistral-7B-v0.1 on the HuggingFaceH4ultrachat_200k dataset. It is the first step (Step 1 SFT, see below) of building Zephyr, i.e. before DPO. It achieves the following results on the evaluation set:

• —Loss: 0.9523

Model description

QLoRA SFT via

# Step 1 - SFT
ACCELERATE_LOG_LEVEL=info accelerate launch --config_file recipes/accelerate_configs/multi_gpu.yaml --num_processes=1 scripts/run_sft.py recipes/zephyr-7b-beta/sft/config_qlora.yaml --load_in_4bit=true

see https://github.com/huggingface/alignment-handbook/blob/main/recipes/zephyr-7b-beta/README.md

Intended uses & limitations

chat_template: "{% for message in messages %}\n{% if message['role'] == 'user' %}\n{{ '<|user|>\n' + message['content'] + eos_token }}\n{% elif message['role'] == 'system' %}\n{{ '<|system|>\n' + message['content'] + eos_token }}\n{% elif message['role'] == 'assistant' %}\n{{ '<|assistant|>\n'  + message['content'] + eos_token }}\n{% endif %}\n{% if loop.last and add_generation_prompt %}\n{{ '<|assistant|>' }}\n{% endif %}\n{% endfor %}"

see https://github.com/huggingface/alignment-handbook/blob/main/recipes/zephyr-7b-beta/sft/config_qlora.yaml

Training and evaluation data

dataset_mixer:
  HuggingFaceH4/ultrachat_200k: 1.0
dataset_splits:
- train_sft
- test_sft

see https://github.com/huggingface/alignment-handbook/blob/main/recipes/zephyr-7b-beta/sft/config_qlora.yaml

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

• —learning_rate: 0.0002
• —trainbatchsize: 1
• —evalbatchsize: 2
• —seed: 42
• —distributed_type: multi-GPU
• —gradientaccumulationsteps: 8
• —totaltrainbatch_size: 8
• —optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• —lrschedulertype: cosine
• —lrschedulerwarmup_ratio: 0.1
• —num_epochs: 1

Training results

Framework versions

• —PEFT 0.7.1
• —Transformers 4.36.2
• —Pytorch 2.1.2
• —Datasets 2.14.6
• —Tokenizers 0.15.0