---
license: mit
language:
- zh
base_model:
- Qwen/Qwen2.5-3B-Instruct
pipeline_tag: text-generation
library_name: transformers
tags:
- finance
- chinese
- qlora
- private-equity
- fund-analysis
- distillation
metrics:
- loss
---

# MachFund-1

A specialized Chinese private equity fund analysis model, fine-tuned from [Qwen2.5-3B-Instruct](https://huggingface.co/Qwen/Qwen2.5-3B-Instruct) using QLoRA knowledge distillation.

## Overview

MachFund-1 is trained to analyze Chinese private equity funds across multiple dimensions: performance analysis, risk assessment, strategy evaluation, manager background, fund comparisons, and investment advice. The model demonstrates a **68.75% improvement** over the base model on domain-specific tasks.

## Training Details

| Parameter | Value |
|---|---|
| Base Model | Qwen2.5-3B-Instruct |
| Method | QLoRA (4-bit NF4 quantization) |
| LoRA Rank / Alpha | 32 / 64 |
| Training Samples | 6,976 (eval: 769) |
| Effective Batch Size | 16 (2 x 8 grad accumulation) |
| Learning Rate | 2e-4 (cosine schedule) |
| Epochs | 2 |
| Max Sequence Length | 6,144 tokens |
| Final Training Loss | 0.9269 |
| Training Time | 141 min on NVIDIA A100 80GB |
| Total Steps | 872 |

### Knowledge Distillation Pipeline

1. **Teacher Model**: Gemini 2.5 Pro generates ~50 Q&A pairs per fund across 8 categories for 178 Chinese private equity funds
2. **Quality Scoring**: Gemini 2.5 Flash scores each pair on 5 dimensions (accuracy, completeness, professionalism, data usage, coherence) with a threshold of 15/25
3. **Student Training**: QLoRA fine-tuning on 6,976 high-quality filtered samples

### Question Categories

- Fund overview and basic information
- Performance analysis and benchmarking
- Risk assessment and drawdown analysis
- Strategy analysis and market positioning
- Manager background and track record
- Fund comparisons (peer and category)
- Investment advice and suitability
- Structured data extraction

## Evaluation

| Gate | Metric | Result |
|---|---|---|
| Training Lift | Base vs Fine-tuned Score | **PASS** (4.8 to 8.1, +68.75%, threshold: 30%) |
| Speed (FP16) | Tokens/sec on RTX 5080 | 30.1 tok/s (threshold: 50) |

## Available Formats

| Format | File | Size | Use Case |
|---|---|---|---|
| SafeTensors (FP16) | `model.safetensors` | 6.17 GB | Full precision inference |
| GGUF Q8_0 | `gguf/mach-fund-1-Q8_0.gguf` | 3.29 GB | High-quality quantized inference |
| GGUF Q4_K_M | `gguf/mach-fund-1-Q4_K_M.gguf` | 1.93 GB | Efficient inference, recommended |
| GGUF F16 | `gguf/mach-fund-1-f16.gguf` | 6.18 GB | Full precision GGUF |

## Usage

### Transformers

```python
from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained("openalchemy/MachFund", torch_dtype="auto", device_map="auto")
tokenizer = AutoTokenizer.from_pretrained("openalchemy/MachFund")

messages = [
    {"role": "system", "content": "You are a professional private equity fund analyst."},
    {"role": "user", "content": "Analyze the performance of this fund"}
]

text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tokenizer(text, return_tensors="pt").to(model.device)
outputs = model.generate(**inputs, max_new_tokens=1024)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))
```

### llama.cpp (GGUF)

```bash
./llama-cli -m mach-fund-1-Q4_K_M.gguf -p "Analyze the risk profile of this fund" -n 512
```

### Ollama

```bash
echo 'FROM ./mach-fund-1-Q4_K_M.gguf' > Modelfile
ollama create machfund -f Modelfile
ollama run machfund "What is the Sharpe ratio of this fund?"
```

## Limitations

- Trained specifically on Chinese private equity fund data; may not generalize to other financial domains
- Training data reflects fund information available up to early 2026
- Should not be used as the sole basis for investment decisions
- Speed on consumer GPUs (RTX 5080) is below the 50 tok/s target at FP16; use GGUF Q4_K_M for faster inference

## License

MIT