---
license: llama3.2
datasets:
- HuggingFaceH4/ultrachat_200k
base_model:
- meta-llama/Llama-3.2-1B
pipeline_tag: text-generation
tags:
- trl
- llama
- sft
- alignment
- transformers
- custome
- chat
---
# Llama-3.2-1B-ultrachat200k


## Model Details

- **Model type:** sft model
- **License:** llama3.2
- **Finetuned from model:** [meta-llama/Llama-3.2-1B](https://huggingface.co/meta-llama/Llama-3.2-1B)
- **Training data:** [HuggingFaceH4/ultrachat_200k](https://huggingface.co/datasets/HuggingFaceH4/ultrachat_200k)
- **Training framework:** [trl](https://github.com/huggingface/trl)

## Training Details

### Training Hyperparameters
`attn_implementation`: flash_attention_2 \
`bf16`: True \
`learning_rate`: 2e-5 \
`lr_scheduler_type`: cosine \
`per_device_train_batch_size`: 2 \
`gradient_accumulation_steps`: 16 \
`torch_dtype`: bfloat16 \
`num_train_epochs`: 1 \
`max_seq_length`: 2048 \
`warmup_ratio`: 0.1

### Results

`init_train_loss`: 1.726 \
`final_train_loss`: 1.22 \

### Training script

```python
import multiprocessing

from datasets import load_dataset
from tqdm.rich import tqdm
from transformers import AutoTokenizer, AutoModelForCausalLM
from trl import (
    ModelConfig,
    SFTTrainer,
    get_peft_config,
    get_quantization_config,
    get_kbit_device_map,
    SFTConfig,
    ScriptArguments,
    TrlParser
)

tqdm.pandas()

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

    quantization_config = get_quantization_config(model_config)
    model_kwargs = dict(
        revision=model_config.model_revision,
        trust_remote_code=model_config.trust_remote_code,
        attn_implementation=model_config.attn_implementation,
        torch_dtype=model_config.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,
                                                 **model_kwargs)
    tokenizer = AutoTokenizer.from_pretrained(
        model_config.model_name_or_path, trust_remote_code=model_config.trust_remote_code, use_fast=True
    )
    tokenizer.pad_token = '<|end_of_text|>'

    train_dataset = load_dataset(args.dataset_name,
                                 split=args.dataset_train_split,
                                 num_proc=multiprocessing.cpu_count())

    trainer = SFTTrainer(
        model=model,
        args=training_args,
        train_dataset=train_dataset,
        processing_class=tokenizer,
        peft_config=get_peft_config(model_config),
    )

    trainer.train()

    trainer.save_model(training_args.output_dir)
```

### Test Script
```python
from vllm import LLM
from datasets import load_dataset
from vllm.sampling_params import SamplingParams
from transformers import AutoTokenizer

MODEL_PATH = "autodl-tmp/saves/Llama-3.2-1B-ultrachat200k"

model = LLM(MODEL_PATH,
            tensor_parallel_size=1,
            dtype='bfloat16')
tokenizer = AutoTokenizer.from_pretrained(MODEL_PATH)

input = tokenizer.apply_chat_template([{"role": "user", "content": "Where is Harbin?"}],
                                    tokenize=False,
                                    add_generation_prompt=True)
sampling_params = SamplingParams(max_tokens=1024,
                                 temperature=0.7,
                                 logprobs=1,
                                 stop_token_ids=[tokenizer.eos_token_id])

vllm_generations = model.generate(input,
                                  sampling_params)

print(vllm_generations[0].outputs[0].text)
# print result: Harbin is located in northeastern China in the Heilongjiang province. It is the capital of Heilongjiang province in the Northeast Asia.
```