Downloads · 30 days
0
elizabeth07-m/email-gym-agent-qwen3-0.6b
email-gym-agent-qwen3-0.6b is a machine learning model from elizabeth07-m. Use it for the machine learning task on the model card, and read the license before you ship it in a product. It is set up for peft. The card lists the license as apache-2.0.
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.
Downloads · 30 days
0
Access
Public
Updated Apr 6, 2026
Repo size
—
Likes
0
Public
Click a slice to open those files.
.md3 KB · 66%
From the Hugging Face model README
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.
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))
| 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 |
| 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 |
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