Tinymistral 248M Chat V4

🚀 TinyMistral-248M-Chat

TinyMistral-248M-Chat is a chat model based on the Locutusque/TinyMistral-248M base model. It is trained on multiple datasets and is licensed under the Apache License 2.0. This model can be used for various text generation tasks, such as answering questions and providing suggestions.

🚀 Quick Start

Prerequisites

• Python environment
• Required Python libraries: transformers, torch

Installation

pip install transformers torch

✨ Features

• Multiple Datasets: Trained on a variety of high - quality datasets, including HuggingFaceH4/ultrachat_200k, Open - Orca/OpenOrca, etc.
• Customizable Prompt Format: Supports a specific prompt format with special tokens <|im_start|> and <|im_end|> for better interaction.
• Fine - Tuned with DPO: Fine - tuned using the DPO method through [LLaMA - Factory](https://github.com/hiyouga/LLaMA - Factory) for better performance.

📦 Installation

pip install transformers torch

💻 Usage Examples

Basic Usage

from transformers import AutoModelForCausalLM, AutoTokenizer, TextStreamer
import torch

model_path = "Felladrin/TinyMistral-248M-Chat-v4"
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
tokenizer = AutoTokenizer.from_pretrained(model_path)
model = AutoModelForCausalLM.from_pretrained(model_path).to(device)
streamer = TextStreamer(tokenizer)
messages = [
    {
        "role": "system",
        "content": "You are a highly knowledgeable and friendly assistant. Your goal is to understand and respond to user inquiries with clarity. Your interactions are always respectful, helpful, and focused on delivering the most accurate information to the user.",
    },
    {
        "role": "user",
        "content": "Hey! Got a question for you!",
    },
    {
        "role": "assistant",
        "content": "Sure! What's it?",
    },
    {
        "role": "user",
        "content": "What are some potential applications for quantum computing?",
    },
]
prompt = tokenizer.apply_chat_template(
    messages, tokenize=False, add_generation_prompt=True
)
inputs = tokenizer(prompt, return_tensors="pt").to(device)
model.generate(
    inputs.input_ids,
    attention_mask=inputs.attention_mask,
    max_length=tokenizer.model_max_length,
    streamer=streamer,
    eos_token_id=tokenizer.eos_token_id,
    pad_token_id=tokenizer.pad_token_id,
    do_sample=True,
    temperature=0.6,
    top_p=0.8,
    top_k=0,
    min_p=0.1,
    typical_p=0.2,
    repetition_penalty=1.176,
)

📚 Documentation

Recommended Prompt Format

<|im_start|>system
{system_message}<|im_end|>
<|im_start|>user
{user_message}<|im_end|>
<|im_start|>assistant

Training Details

SFT Training

This model was trained with SFTTrainer using the following settings:

DPO Fine - Tuning

Then, the model was fine - tuned with DPO through [LLaMA - Factory](https://github.com/hiyouga/LLaMA - Factory) using the following hyperparameters and command:

llamafactory-cli train \
    --stage dpo \
    --do_train True \
    --model_name_or_path ~/TinyMistral-248M-Chat \
    --preprocessing_num_workers $(python -c "import os; print(max(1, os.cpu_count() - 2))") \
    --dataloader_num_workers $(python -c "import os; print(max(1, os.cpu_count() - 2))") \
    --finetuning_type full \
    --flash_attn auto \
    --enable_liger_kernel True \
    --dataset_dir data \
    --dataset ultrafeedback \
    --cutoff_len 1024 \
    --learning_rate 1e-6 \
    --num_train_epochs 2.0 \
    --per_device_train_batch_size 4 \
    --gradient_accumulation_steps 4 \
    --lr_scheduler_type linear \
    --max_grad_norm 1.0 \
    --logging_steps 10 \
    --save_steps 50 \
    --save_total_limit 1 \
    --warmup_ratio 0.1 \
    --packing False \
    --report_to tensorboard \
    --output_dir ~/TinyMistral-248M-Chat-v4 \
    --pure_bf16 True \
    --plot_loss True \
    --trust_remote_code True \
    --ddp_timeout 180000000 \
    --include_tokens_per_second True \
    --include_num_input_tokens_seen True \
    --optim adamw_8bit \
    --pref_beta 0.5 \
    --pref_ftx 0 \
    --pref_loss simpo \
    --gradient_checkpointing True

📄 License

This project is licensed under the Apache License 2.0.