Qwen2.5-1.5B-SimPO

Model Details

• Model type: aligned model
• License: Apache license 2.0
• Finetuned from model: AIR-hl/Qwen2.5-1.5B-ultrachat200k
• Training data: HuggingFaceH4/ultrafeedback_binarized
• Training framework: trl

Training Details

devices: 4 * NPU 910B-64GB
precision: bf16 mixed-precision
global_batch_size: 128

Training Hyperparameters

beta: 1
gamma: 0.1
bf16: True
learning_rate: 1e-6
lr_scheduler_type: cosine
per_device_train_batch_size: 16
gradient_accumulation_steps: 2
torch_dtype: bfloat16
num_train_epochs: 1
max_prompt_length: 512
max_length: 1024
warmup_ratio: 0.05

Results

init_train_loss: 0.7551
final_train_loss: 0.6715
accuracy: 0.6375
reward_margin: 0.3633

Training script

import torch
from datasets import load_dataset
from transformers import AutoModelForCausalLM, AutoTokenizer
from trl import (
    CPOConfig,
    CPOTrainer,
    ModelConfig,
    ScriptArguments,
    TrlParser,
    get_kbit_device_map,
    get_peft_config,
    get_quantization_config,
)
from trl.trainer.utils import SIMPLE_CHAT_TEMPLATE

if __name__ == "__main__":
    parser = TrlParser((ScriptArguments, CPOConfig, ModelConfig))
    script_args, training_args, model_config = parser.parse_args_and_config()

    torch_dtype = (
        model_config.torch_dtype
        if model_config.torch_dtype in ["auto", None]
        else getattr(torch, model_config.torch_dtype)
    )

    quantization_config = get_quantization_config(model_config)

    model_kwargs = dict(
        revision=model_config.model_revision,
        attn_implementation=model_config.attn_implementation,
        torch_dtype=torch_dtype,
        use_cache=False if training_args.gradient_checkpointing else True,
        device_map=get_kbit_device_map() if quantization_config is not None else None,
        quantization_config=quantization_config,
    )

    model = AutoModelForCausalLM.from_pretrained(
        model_config.model_name_or_path, trust_remote_code=model_config.trust_remote_code, **model_kwargs
    )

    peft_config = get_peft_config(model_config)

    tokenizer = AutoTokenizer.from_pretrained(
        model_config.model_name_or_path, trust_remote_code=model_config.trust_remote_code
    )
    if tokenizer.pad_token is None:
        tokenizer.pad_token = tokenizer.eos_token
    if tokenizer.chat_template is None:
        tokenizer.chat_template = SIMPLE_CHAT_TEMPLATE
    if script_args.ignore_bias_buffers:
        model._ddp_params_and_buffers_to_ignore = [
            name for name, buffer in model.named_buffers() if buffer.dtype == torch.bool
        ]

    dataset=load_dataset(script_args.dataset_name,
                         split=script_args.dataset_train_split)
    dataset=dataset.select_columns(['prompt', 'chosen', 'rejected'])

    trainer = CPOTrainer(
        model,
        args=training_args,
        train_dataset=dataset,
        processing_class=tokenizer,
        peft_config=peft_config,
    )

    trainer.train()

    trainer.save_model(training_args.output_dir)