jinv2/qwen-1.8b-chat-lora-stock-quant-edu

Qwen-1.8B-Chat LoRA for Stock Market Quantitative Education (股票量化投教LoRA模型)

This repository contains LoRA (Low-Rank Adaptation) adapters fine-tuned on the Qwen/Qwen-1_8B-Chat model. The goal of this fine-tuning is to create an AI assistant that can explain stock market and quantitative trading concepts in plain language ("大白话"), making these topics more accessible to beginners.

Model Description

This model is a PEFT-LoRA adaptation of the Qwen/Qwen-1_8B-Chat large language model. It has been fine-tuned on a small, custom dataset of ~20 instruction-response pairs focused on financial education. Due to the small dataset size, this model should be considered experimental and for demonstration purposes.

Developed by: 天算AI科技研发实验室 (Natural Algorithm AI R&D Lab) - jinv2

Intended Uses & Limitations

Intended Uses:

• Educational tool for understanding basic stock market and quantitative trading terms.
• Generating simple explanations of financial concepts.
• Demonstrating the LoRA fine-tuning process on a chat model for a specific domain.

Limitations:

• Not for Financial Advice: The information provided by this model is strictly for educational purposes and should NOT be considered financial advice. Always consult with a qualified financial advisor before making investment decisions.
• Limited Knowledge: Fine-tuned on a very small dataset. Its knowledge is restricted and may not be comprehensive or entirely accurate.
• Potential for Hallucinations: Like all LLMs, it may generate incorrect or nonsensical information.
• Overfitting: Due to the small dataset, the model may be overfit to the training examples.
• Bias: The training data may contain biases, which could be reflected in the model's responses.
• Requires Base Model: These are LoRA adapters and require the original Qwen/Qwen-1_8B-Chat base model to be loaded first.

How to Use with PEFT

You would typically load the base model and then apply these LoRA adapters using the PEFT library.

from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

base_model_name = "Qwen/Qwen-1_8B-Chat"
adapter_model_name = "jinv2/qwen-1.8b-chat-lora-stock-quant-edu" # Replace with your actual model name on Hub

# Load tokenizer
tokenizer = AutoTokenizer.from_pretrained(base_model_name, trust_remote_code=True)
if tokenizer.pad_token_id is None:
    tokenizer.pad_token_id = tokenizer.eos_token_id # Or <|endoftext|> ID: 151643

# Load base model
base_model = AutoModelForCausalLM.from_pretrained(
    base_model_name,
    torch_dtype=torch.float16, # or "auto"
    device_map="auto",
    trust_remote_code=True
)

# Load LoRA adapter
model = PeftModel.from_pretrained(base_model, adapter_model_name)
model = model.eval() # Set to evaluation mode

# Example Inference (Qwen chat format)
prompt = "请用大白话解释什么是MACD指标？"
# For Qwen-Chat, using model.chat() is recommended
response, history = model.chat(tokenizer, prompt, history=None, system="You are a helpful financial education assistant.")
print(response)

# Alternative generic generation
# messages = [
#     {"role": "system", "content": "You are a helpful financial education 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(model.device)
# generated_ids = model.generate(model_inputs.input_ids, max_new_tokens=512)
# 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

Base Model: Qwen/Qwen-1_8B-Chat

Fine-tuning Method: LoRA (Low-Rank Adaptation) via PEFT and TRL SFTTrainer.

Dataset: ~20 custom instruction-response pairs for financial education.

Training Configuration (Key Parameters):

LoRA r: 8

LoRA alpha: 16

Target Modules: c_attn, c_proj, w1, w2

Optimizer: AdamW (default from Trainer)

Precision: FP32 (due to issues with FP16/BF16 GradScaler in the environment)

Epochs: ~17 (based on 80 steps)

Batch Size (effective): 4 (per_device_train_batch_size=1, gradient_accumulation_steps=4)

Learning Rate: 2e-4

Max Sequence Length: 512

Disclaimer

This model is provided "as-is" without any warranty. The developers are not responsible for any outcomes resulting from the use of this model. Always verify information and use at your own risk.

Copyright Information:

© 天算AI科技研发实验室 (Natural Algorithm AI R&D Lab) - jinv2
All rights reserved unless otherwise specified by the license.