firefly-qwen1.5-en-7b-unsloth

Unsloth x Qwen2

Unsloth can speed up training LLM and reduce memory usage, but currently it only supports Llama3, Mistral, Gemma, ORPR, Phi-3 and TinyLlama. We can't train Qwen2 with Unsloth, even though Qwen2 is popular in community.

It's exciting that we succeed to make Unsloth support Qwen2, it can speed up training and reduce much memory usage. If you want to train Qwen2 with Unsloth, you can use our repo rather than the official one. And we will commit our code to the official repo.

Install our Unsloth:

pip install git+https://github.com/yangjianxin1/unsloth.git

Firefly already supports training Qwen2 with Unsloth, and the subsequent models are trained with Firefly, you can try it.

Model Card for Firefly-Qwen1.5-Unsloth

firefly-qwen1.5-en-7b-unsloth and firefly-qwen1.5-en-7b-dpo-v0.1-unloth are trained based on Qwen1.5-7B to act as a helpful and harmless AI assistant. We use Firefly to train our models on a single V100 GPU with QLoRA and Unsloth. firefly-qwen1.5-en-7b-unsloth is fine-tuned based on Qwen1.5-7B with English instruction data, and firefly-qwen1.5-en-7b-dpo-v0.1-unsloth is trained with Direct Preference Optimization (DPO) based on firefly-qwen1.5-en-7b-unsloth.

Our models outperform official Qwen1.5-7B-Chat, Gemma-7B-it, Zephyr-7B-Beta on Open LLM Leaderboard.

Although our models are trained with English data, you can also try to chat with models in Chinese because Qwen1.5 is also good at Chinese. But we have not evaluated the performance in Chinese yet.

We advise you to install transformers>=4.37.0.

Performance

We have evaluated the training gain of Qwen1.5-7B, we use QLoRA and Unsloth to train model for 20 steps on a single V100. The result can be listed as follows. Unsloth can reduce GPU memory by 39.13% and training time by 32.12%, and the training speed can increase by 47.32%.

We evaluate our sft and dpo models on Open LLM Leaderboard, they achieve good performance.

Usage

The chat templates of our chat models are the same as Official Qwen1.5-7B-Chat:

<|im_start|>system
You are a helpful assistant.<|im_end|>
<|im_start|>user
hello, who are you?<|im_end|>
<|im_start|>assistant
I am a AI program developed by Firefly<|im_end|>

You can use script to inference in Firefly.

You can also use the following code:

from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

model_name_or_path = "YeungNLP/firefly-qwen1.5-en-7b-unsloth"
model = AutoModelForCausalLM.from_pretrained(
    model_name_or_path,
    trust_remote_code=True,
    low_cpu_mem_usage=True,
    torch_dtype=torch.float16,
    device_map='auto',
)
tokenizer = AutoTokenizer.from_pretrained(model_name_or_path)

prompt = "Compose an engaging travel blog post about a recent trip to Hawaii, highlighting cultural experiences and must-see attractions. "
messages = [
    {"role": "system", "content": "You are a helpful assistant."},
    {"role": "user", "content": prompt}
]
text = tokenizer.apply_chat_template(
    messages,
    tokenize=False,
    add_generation_prompt=True
)
model_inputs = tokenizer([text], return_tensors="pt").to('cuda')

generated_ids = model.generate(
    model_inputs.input_ids,
    max_new_tokens=1500,
    top_p = 0.9,
    temperature = 0.35,
    repetition_penalty = 1.0,
    eos_token_id=tokenizer.encode('<|im_end|>', add_special_tokens=False)
)
generated_ids = [
    output_ids[len(input_ids):] for input_ids, output_ids in zip(model_inputs.input_ids, generated_ids)
]

response = tokenizer.batch_decode(generated_ids, skip_special_tokens=True)[0]
print(response)

Training Details

Both in SFT and DPO stages, We only use a single V100 GPU with QLoRA and Unsloth, and we use Firefly to train our models.

Training Setting

The following hyperparameters are used during SFT:

• num_epochs: 1
• learning_rate: 2e-4
• total_train_batch_size: 32
• max_seq_length: 2048
• optimizer: paged_adamw_32bit
• lr_scheduler_type: constant_with_warmup
• warmup_steps: 600
• lora_rank: 64
• lora_alpha: 16
• lora_dropout: 0.05
• gradient_checkpointing: true
• fp16: true

The following hyperparameters were used during DPO:

• num_epochs: 1
• learning_rate: 2e-4
• total_train_batch_size: 32
• max_seq_length: 2048
• max_prompt_length: 500
• optimizer: paged_adamw_32bit
• lr_scheduler_type: constant_with_warmup
• warmup_steps: 100
• lora_rank: 64
• lora_alpha: 16
• lora_dropout: 0.05
• gradient_checkpointing: true
• fp16: true

Training metrics

The table below shows the full set of DPO training metrics: