elizabeth07-m/email-gym-agent-qwen3-0.6b

🧠 Message Routing Agent — GRPO Fine-tuned Qwen3-0.6B

A GRPO-fine-tuned LoRA adapter for automated message triage, routing, and operational response generation. This model achieves high accuracy acting as an autonomous triage assistant in the OpenEnv message-routing-gym.

Quick Start

from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer

# Load base model + LoRA adapter
base = AutoModelForCausalLM.from_pretrained("Qwen/Qwen3-0.6B", torch_dtype="auto", trust_remote_code=True)
model = PeftModel.from_pretrained(base, "elizabeth07-m/email-gym-agent-qwen3-0.6b")
tokenizer = AutoTokenizer.from_pretrained("Qwen/Qwen3-0.6B", trust_remote_code=True)

messages = [
    {"role": "system", "content": "You are an automated message routing specialist. Given a DIRECTIVE, SOURCE, and MESSAGE, output JSON with action_type, message_id, target_directory, and response_payload."},
    {"role": "user", "content": "DIRECTIVE: Route P1s to critical.\nSOURCE: ops-pager\nMESSAGE: ID 1: Database down."}
]
text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True, enable_thinking=False)
inputs = tokenizer(text, return_tensors="pt").to(model.device)
output = model.generate(**inputs, max_new_tokens=256)
print(tokenizer.decode(output[0][inputs["input_ids"].shape[1]:], skip_special_tokens=True))

Model Details

Base model	Qwen/Qwen3-0.6B
License	Apache 2.0 (ungated)
Method	GRPO (Group Relative Policy Optimization)
LoRA	rank=16, alpha=32
Training	3 epochs, lr=5e-6, beta=0.04, 2 generations/prompt

Results

Task	Baseline	GRPO	Δ
Easy: Single routing	0.4000	0.9000	+0.5000
Medium: Adversarial noise	0.3500	0.7800	+0.4300
Hard: Multi-step response	0.2000	0.6500	+0.4500
Average	0.3166	0.7766	+0.4600

Reproduce

git clone https://github.com/elizabeth07-m/email_gym
cd email_gym
# Install dependencies
uv sync
# Run training notebook
# Open notebooks/email_gym_grpo_training.ipynb and run all cells

Framework

• TRL + PEFT + Transformers
• Trained for OpenEnv Benchmarking on RTX 4090