entity-extraction-smollm

Entity Extraction Model - Fine-tuned SmolLM-360M

This model is a fine-tuned version of HuggingFaceTB/SmolLM-360M for extracting structured entities from natural language calendar event descriptions.

Model Description

• Base Model: HuggingFaceTB/SmolLM-360M
• Task: Entity Extraction for Calendar Events
• Language: English
• License: Apache 2.0

Intended Use

This model extracts structured entities from natural language text describing calendar events. It outputs JSON with the following fields:

• action: Type of event (e.g., "meeting", "lunch")
• date: Date in DD/MM/YYYY format
• time: Time in HH:MM AM/PM format
• attendees: Array of attendee names (or null)
• location: Event location (or null)
• duration: Duration description (or null)
• recurrence: Recurrence pattern (or null)
• notes: Additional notes (or null)

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

# Load model and tokenizer
model_name = "Shresth12345/entity-extraction-smollm"
model = AutoModelForCausalLM.from_pretrained(model_name)
tokenizer = AutoTokenizer.from_pretrained(model_name)

# Example usage
text = "Meeting with John tomorrow at 2pm for 1 hour at the office"
prompt = f"Extract entities from: {text}"

# Tokenize and generate
inputs = tokenizer(prompt, return_tensors="pt", truncation=True, max_length=512)
with torch.no_grad():
    outputs = model.generate(
        **inputs,
        max_new_tokens=150,
        temperature=0.1,
        do_sample=True,
        pad_token_id=tokenizer.eos_token_id
    )

# Decode response
response = tokenizer.decode(outputs[0], skip_special_tokens=True)
generated_text = response[len(prompt):].strip()
print(generated_text)

Expected Output Format

{
  "action": "Meeting",
  "date": "tomorrow", 
  "time": "2:00 PM",
  "attendees": ["John"],
  "location": "office",
  "duration": "1 hour",
  "recurrence": null,
  "notes": null
}

Training Details

• Training Data: 793 calendar event samples
• Training Split: 70% train, 15% validation, 15% test
• Custom Loss Function: Entity-aware loss with weighted output portion
• Training Framework: PyTorch (custom trainer)
• Evaluation Metrics: Exact match accuracy, field-wise accuracy, JSON quality

Model Performance

The model demonstrates strong performance in:

• Accurate entity extraction from natural language
• Consistent JSON output format
• Handling of missing/null values
• Recognition of temporal expressions
• Identification of people and locations

Limitations

• Primarily trained on English calendar events
• May struggle with very complex or ambiguous temporal expressions
• Performance may vary with domain-specific terminology
• Requires specific input format: "Extract entities from: [text]"

Training Procedure

This model was fine-tuned using:

1. Custom PyTorch trainer implementation
2. Entity-weighted loss function (weight: 2.0)
3. Cosine annealing learning rate schedule
4. Gradient accumulation for effective larger batch sizes
5. Early stopping based on validation performance

Citation

If you use this model, please cite:

@misc{entity-extraction-smollm,
  title={Entity Extraction Fine-tuned SmolLM-360M},
  author={Shresth Mishra},
  year={2025},
  howpublished={\url{https://huggingface.co/Shresth12345/entity-extraction-smollm}}
}

Contact

For questions about this model, please open an issue in the repository or contact the author.