WKQ9411/Mini-Llama3-100M-SFT

Mini-LLM

Mini-LLM is a project that aims to replicate mainstream open-source model architectures with limited computational resources, implementing mini models with 100-200M parameters. The project focuses on learning and reproducing model architectures while providing complete training and inference pipelines. For more details, please visit the Mini-LLM project.

Usage

Using Transformers Library

First, import the model registration module, then load the model using AutoModelForCausalLM:

import mini_models  # Register custom Mini-LLM models
from transformers import AutoTokenizer, AutoModelForCausalLM
import torch

model = AutoModelForCausalLM.from_pretrained("WKQ9411/Mini-Llama3-100M-SFT")
tokenizer = AutoTokenizer.from_pretrained("WKQ9411/Mini-Llama3-100M-SFT")
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
model = model.to(device)

# Generate text
messages = [{"role": "user", "content": "你好，你是谁？"}]
formatted_text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
input_ids = tokenizer(formatted_text, return_tensors="pt")["input_ids"].to(model.device)
response = model.generate(input_ids, max_new_tokens=100)
response = tokenizer.decode(response[0][len(input_ids[0]):], skip_special_tokens=True)
print(response)

Using Custom Interface

from mini_models import get_model_and_config
from transformers import AutoTokenizer
import torch

Model, Config = get_model_and_config("mini_llama3")
model = Model.from_pretrained("path/to/your/model")
tokenizer = AutoTokenizer.from_pretrained("path/to/your/tokenizer")
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
model = model.to(device)

# Use the model for generation
messages = [{"role": "user", "content": "你好，你是谁？"}]
formatted_text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
input_ids = tokenizer(formatted_text, return_tensors="pt")["input_ids"].to(model.device)
response = model.generate(input_ids, max_new_tokens=100)
response = tokenizer.decode(response[0][len(input_ids[0]):], skip_special_tokens=True)
print(response)

Training Data

The model was pre-trained on:

• 20% sampled subset of OpenCSG Fineweb-Edu-Chinese-V2.1 dataset (high-quality Chinese educational content)
• DeepCtrl SFT dataset from Modelscope

Limitations

This is a small-scale model designed for educational and research purposes. It may not perform as well as larger models on complex tasks.