abababab2003/trader-sft-lora

ActiveTrader — SFT Fine-Tuned Trader Agent

LoRA adapter for Qwen3-8B, fine-tuned to generate structured stock trading recommendations from analyst and risk manager reports.

What This Model Does

Takes two inputs from upstream agents:

1. Analyst Report — fundamentals, news, social sentiment, macro context
2. Risk Manager Report — technical indicators, support/resistance, risk assessment

Outputs a structured Trading Recommendation: Buy / Hold / Sell with entry zone, stop loss, target price, reasoning, and key risks.

Training Details

Parameter	Value
Base model	Qwen/Qwen3-8B
Method	QLoRA (4-bit NF4)
LoRA rank	16
LoRA alpha	32
Target modules	q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj
Training examples	150 (30 tickers × 5 question variants)
Train/eval split	135 / 15
Epochs	3
Batch size	2 × 4 grad accum = 8 effective
Learning rate	2e-4 (cosine schedule)
Hardware	NVIDIA RTX 4070 (8GB VRAM)
Training time	~9 hours
Trainable params	43.6M / 8.2B (0.53%)

Training Results

Metric	Value
Initial train loss	1.845
Final train loss	0.481
Final eval loss	0.534

Training Data

150 SFT examples generated by:

1. Running Analyst (Qwen2.5-7B) + Risk Manager (Qwen2.5-7B) on 30 tickers across sectors (tech, finance, healthcare, energy, consumer, industrial)
2. Sending report pairs to GPT-4o with varied user questions to generate gold-standard trader recommendations
3. Formatting as chat-style JSONL (system + user + assistant)

Usage

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

# Load base model in 4-bit
base_model = AutoModelForCausalLM.from_pretrained(
    "Qwen/Qwen3-8B",
    quantization_config=BitsAndBytesConfig(
        load_in_4bit=True,
        bnb_4bit_compute_dtype=torch.bfloat16,
        bnb_4bit_quant_type="nf4",
    ),
    device_map="cuda:0",
)

# Load LoRA adapter
model = PeftModel.from_pretrained(base_model, "abababab2003/trader-sft-lora")
model = model.merge_and_unload()

tokenizer = AutoTokenizer.from_pretrained("Qwen/Qwen3-8B")

Project

ActiveTrader — a multi-agent trading system built with LangGraph for CS 496 (Agent AI) at Northwestern University. Three agents collaborate: an Analyst, a Risk Manager, and this SFT-trained Trader.

• GitHub: Vio1etV/Trading-Agent

Framework Versions

• PEFT: 0.17.1
• Transformers: 4.57.6
• PyTorch: 2.6.0+cu124
• bitsandbytes: 0.48.2