Qwen3-30B-A3B-Instruct-Fund-Rating

This model is a fine-tuned version of Qwen/Qwen3-30B-A3B-Instruct-2507 on the fund_set_cn dataset. It achieved the following results on the evaluation set:

Loss: 0.3771

Model Description

This model is fine-tuned based on Qwen/Qwen3-30B-A3B-Instruct-2507 using a proprietary fund dataset containing fund introductions and Morningstar ratings. The dataset includes data from over 10,000 Chinese funds. The training employed techniques such as QLoRA and flash attention, and was conducted on NVIDIA A100-PCIE-40GB. The final model can provide AI-based ratings for funds based on their introductions, with an average difference from Morningstar ratings of approximately 0.5.

Usage Scope and Limitations

Suitable for rating new funds. Ratings include: 5-star, 4-star, 3-star, 2-star, 1-star. 5-star is the highest rating, 1-star is the lowest rating.

import os
from openai import OpenAI

# Configure client parameters
client = OpenAI(
    base_url="http://localhost:8000/v1",  # Address and port of the vLLM API server
    api_key="token-abc123"  # If the vLLM service does not require authentication, any string can be filled in
)

def chat_with_model():
    
    messages = [
        {"role": "user", "content": "Please rate the following fund and provide the rationale for the rating. Fund Introduction: China Universal Enhanced Bond Fund A (Code 519078) is a Bond-Mixed Primary fund managed by China Universal Fund Management Company. The current fund manager, Gan Xinyu, has been in position since June 2023. The fund was established in March 2008, with a current asset size of 3.368 billion CNY. Its investment objective is to primarily invest in fixed-income instruments such as bonds and low-risk assets while strictly controlling risks, aiming for stable returns. Investment strategies include category allocation, money market instrument management, and IPO subscriptions, using the ChinaBond Total Index as the performance benchmark. Historical performance has been steady, with a 1-year return of 5.17%, a 3-year return of 11.91%, and a cumulative return since inception of 111.93%. Risk control is excellent, with a 1-year annualized volatility of only 2.02%, a maximum drawdown of 1.3%, and a Sharpe ratio of 1.82. Both risk-return ratios and anti-risk volatility indicators outperform 49% of peer funds. The fund has an active dividend policy, distributing up to 12 times per year, with Class A shares distributing no less than 25% of the annual distributable profits. Fund manager Gan Xinyu has extensive fixed-income research experience and manages multiple bond funds, demonstrating professional management capabilities."},
    ]

    # Call the model
    response = client.chat.completions.create(
        model="/root/autodl-tmp/models/Qwen3-30B-A3B-Instruct-Fund-Rating",  # Your deployed model path
        messages=messages,
        temperature=0.7,  # Controls the randomness of responses, range 0-1, higher values make responses more creative
        max_tokens=2048,   # Controls the maximum length of the response
        stream=False       # Set to True to enable streaming output
    )
    
    # Get the model's reply
    assistant_reply = response.choices[0].message.content
    
    # Print the assistant's reply
    print(f"\n🤖 Assistant: {assistant_reply}")

if __name__ == "__main__":
    chat_with_model()

Training and Evaluation Data

The training and validation data are structured as follows:

[
    {
        "instruction": "Please provide an AI rating for this fund based on its profile. Ratings include: 5-star, 4-star, 3-star, 2-star, and 1-star, with 5-star being the highest and 1-star the lowest.",
        "input": "Fund Profile: Huashang Shengshi Chengzhang Hybrid (Code 630002) is a hybrid-equity fund established in 2008, jointly managed by Sun Wei and Wang Yiwen. The fund has an asset size of 3.12 billion CNY, with a high equity allocation of 88.87%. It primarily invests in high-growth sectors such as manufacturing (51.03%), mining (13.97%), and information technology (6.22%). The fund employs a combination of top-down and bottom-up strategies, focusing on companies with sustainable growth potential and reasonable valuations to benefit from China's rapid economic growth. It has demonstrated outstanding historical performance, with a one-year return of 58.86%, a three-year return of 46.7%, and a cumulative return of over 1000% since inception. In terms of risk control, the one-year annualized Sharpe ratio is 2.44, and the maximum drawdown is 12.21%, with a risk-return ratio outperforming 78% of its peers, reflecting strong profitability and risk-adjusted returns.",
        "output": "Fund Rating: 5-star, Rating Rationale: Exceptional performance with a one-year return of 58.86%, excellent risk-return ratio, Sharpe ratio of 2.44, well-controlled maximum drawdown, and a clear investment strategy focused on high-growth sectors."
    },
    {
        "instruction": "Please provide an AI rating for this fund based on its profile. Ratings include: 5-star, 4-star, 3-star, 2-star, and 1-star, with 5-star being the highest and 1-star the lowest.",
        "input": "Fund Profile: Guotai Quantitative Strategy Income Hybrid A (Code 000199) is a hybrid-equity fund managed by Guotai Fund, established in August 2013. The fund size is 160 million CNY, employing a quantitative multi-factor stock selection strategy combined with fundamental research to select individual stocks. The equity allocation in the portfolio is 82.17%, primarily concentrated in manufacturing (56.4%) and financial sectors (8.44%). Fund managers Gao Chongnan and He Tianyuan have extensive quantitative investment experience, achieving a return of 29.04% during their tenure. The fund has performed exceptionally well over the past year, with a return of 53.74%, excellent risk control, an annualized Sharpe ratio of 2.54, and a maximum drawdown of only 10.89%, ranking in the top 80% and 87% for risk-return ratio and anti-risk volatility among peers, respectively. The fund uses tools such as stock index futures to hedge risks, aiming to exceed the performance benchmark (75% CSI 300 Index return + 25% CSI Composite Bond Index return).",
        "output": "Fund Rating: 4-star, Rating Rationale: The quantitative strategy performs excellently, with a standout one-year return of 53.74% and superior risk control (maximum drawdown of 10.89%). However, the small fund size and higher long-term volatility warrant a 4-star rating."
    },
    {
        "instruction": "Please provide an AI rating for this fund based on its profile. Ratings include: 5-star, 4-star, 3-star, and 1-star, with 5-star being the highest and 1-star the lowest.",
        "input": "Fund Profile: Ping An Yuanhe 90-Day Rolling Short-Term Bond A (Code 014468) is a hybrid primary bond fund under Ping An Fund, managed by fund manager Liu Xiaolan. The fund was established on May 18, 2022, with a current asset size of 1.999 billion CNY. Its investment objective is to pursue returns exceeding the performance benchmark while strictly controlling risks and maintaining high liquidity. The fund primarily invests in fixed-income instruments such as bonds, asset-backed securities, and interbank certificates of deposit. It does not directly invest in stocks but may hold shares acquired through convertible bond conversions, which must be sold within 10 trading days. Investment strategies include portfolio duration allocation, asset category allocation, and individual bond selection to manage interest rate and credit risks. Historical performance shows a cumulative return of 10.05% since inception, a one-year return of 1.85%, and a three-year return of 9.05%. In terms of risk, the one-year annualized volatility is 0.41%, with a maximum drawdown of 0.41% and a Sharpe ratio of 0.86. Over three years, the Sharpe ratio improves to 3.3, with strong anti-risk volatility (ranking in the top 97% among peers). However, the risk-return ratio is relatively low (ranking 30% among peers over one year). The fund has low fees, with a management fee of 0.20% and a custody fee of 0.05%, making it suitable for investors with low to moderate risk tolerance.",
        "output": "Fund Rating: 3-star, Rating Rationale: The fund demonstrates good risk control, low volatility, and drawdown, with a high Sharpe ratio over three years. However, its one-year risk-return ratio is relatively low (30% among peers), and its performance is steady but not outstanding, making it suitable for investors with moderate risk tolerance."
    },
    ...
]

Training Procedure

1. Prepare basic fund information (fund code, name, manager, company, investment strategy, etc., along with historical performance data).
2. Use the Deepseek API to extract key data and convert it into a question-answer format.
3. Install necessary dependencies and use LLaMA Factory for training.
4. Merge the base model and adapter.
5. Test and validate the fine-tuned model.

Training Hyperparameters

The following hyperparameters were used during training:

• learning_rate: 5e-05
• train_batch_size: 4
• eval_batch_size: 2
• seed: 42
• gradient_accumulation_steps: 4
• total_train_batch_size: 16
• optimizer: Use OptimizerNames.ADAMW_TORCH with betas=(0.9,0.999) and epsilon=1e-08 and optimizer_args=No additional optimizer arguments
• lr_scheduler_type: cosine
• lr_scheduler_warmup_ratio: 0.1
• num_epochs: 2.0

Training Results

Training Loss	Epoch	Step	Validation Loss
0.5005	0.3331	200	0.4973
0.4418	0.6661	400	0.4335
0.3894	0.9992	600	0.4048
0.3462	1.3314	800	0.3878
0.3493	1.6644	1000	0.3792
0.3477	1.9975	1200	0.3768

Framework Versions

• PEFT 0.15.2
• Transformers 4.52.4
• Pytorch 2.8.0+cu128
• Datasets 3.6.0
• Tokenizers 0.21.1