Shanvit/email-triage-slm

Email Triage SLM

LoRA fine-tuned small language model for email triage. Classifies incoming emails and generates structured triage outputs including classification, priority, suggested actions, and draft responses.

Model Details

• Base Model: Qwen/Qwen3-4B-Instruct-2507 (4B parameters)
• Fine-tuning Method: LoRA (Low-Rank Adaptation)
• Task: Email classification and structured triage
• License: MIT

Training Details

Hyperparameters

Parameter	Value
LoRA Rank	32
Learning Rate	1e-4
Batch Size	4
Epochs	3
Loss	Cross-entropy (prompt masked, completion weighted)

Training Data

• Dataset: jason23322/high-accuracy-email-classifier
• Categories: Spam, Promotions, Verify_code, Updates, Forum, Social_media
• Format: Structured prompt/completion pairs with EOS tokens for stopping behavior

Output Format

The model produces structured triage output in four steps:

Step 1: Classification → {category}
Step 2: Priority → {High|Medium|Low}
Step 3: Suggested actions → {action_items}
Step 4: Drafted response → {draft_response}

Category-to-Priority Mapping:

• verify_code → High priority, urgent action required
• spam / promotions → Low priority, delete/mark as spam
• updates → Medium priority, read and acknowledge
• forum / social_media → Medium priority, read later

Usage

With Transformers + PEFT

from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel

base_model = "Qwen/Qwen3-4B-Instruct-2507"
adapter_path = "path/to/adapters/email_triage_final"

model = AutoModelForCausalLM.from_pretrained(base_model)
model = PeftModel.from_pretrained(model, adapter_path)
tokenizer = AutoTokenizer.from_pretrained(base_model)

prompt = """You are an efficient admin assistant. Analyze this incoming email step-by-step and output in structured format:

Email:
Subject: Your order has shipped!
Hi [Customer Name], Your recent order #[ORDER NUMBER] has shipped and is on its way. You can track your package here: [Tracking Link]. Thanks for shopping with us!"""

inputs = tokenizer(prompt, return_tensors="pt")
outputs = model.generate(**inputs, max_new_tokens=256, temperature=0.7, do_sample=True)
response = tokenizer.decode(outputs[0], skip_special_tokens=True)

Inference API

When running the included FastAPI server:

curl -X POST http://localhost:8000/inference \
  -H "Content-Type: application/json" \
  -d '{
    "prompt": "You are an efficient admin assistant. Analyze this incoming email step-by-step and output in structured format:\n\nEmail:\nSubject: Your order has shipped!\n...",
    "max_tokens": 256,
    "temperature": 0.7
  }'

Important: Use the exact prompt format above for best results. The model was trained with this structure.

Intended Use

• Email inbox triage and prioritization
• Administrative assistant workflows
• Automated email classification and routing
• Draft response generation for common email types

Limitations

• Trained on a specific dataset (400 examples default); may not generalize to all email domains
• Structured output format is fixed; customization requires retraining
• Base model (Qwen3-4B) is hardcoded; adapter is not compatible with other base models
• May misclassify edge cases or emails in languages underrepresented in training data

Bias & Safety

• Inherits base model biases; evaluate outputs for your use case
• Training data reflects the distribution of the source dataset
• Not intended for high-stakes decisions without human review

Training Infrastructure

Trained using Tinker Labs distributed training API. For full training pipeline, dataset preparation, and local inference setup, see the project repository.

Citation

@misc{email-triage-slm,
  title={Email Triage SLM: LoRA Fine-tuned Model for Email Classification},
  author={Shanvit S Shetty},
  year={2026},
  url={https://huggingface.co/your-username/email-triage-slm}
  howpublished={\url{https://github.com/Shanvit7/email-triage-slm}}
}

Repository: github.com/Shanvit7/email-triage-slm