cyqwill/sger-lfm2.5-name-matching

TEXT GENERATIONPricing:Input $0.04 / Cached $0.008 / Output $0.08Concurrent Unit Cost:1Model Size:0.35BQuant:BF16Context Size:32kPublished:Aug 2, 2026License:otherArchitecture:Transformer Featherless Exclusive Cold

The cyqwill/sger-lfm2.5-name-matching model is a 0.35 billion parameter language model, fine-tuned from LiquidAI/LFM2.5-350M, specifically designed for Indian KYC (Know Your Customer) scenarios. It excels at two-stage name processing: first, parsing noisy names into first, middle, and last names (including Devanagari script), and second, performing binary name matching to determine if two names refer to the same person. This model is optimized for handling variations like spelling differences, abbreviations, token reordering, merged tokens, and honorifics in Indian names, achieving high precision and recall on relevant datasets.

Loading preview...

Overview

The cyqwill/sger-lfm2.5-name-matching model is a specialized 0.35 billion parameter language model, built upon the LiquidAI/LFM2.5-350M base, and fine-tuned through a two-stage LoRA process. Its primary application is in Indian KYC (Know Your Customer) scenarios, focusing on robust name processing and matching.

Key Capabilities

  • Noisy Name Parsing: Accurately extracts first_name, middle_name, and last_name from noisy input names, supporting Devanagari (天城文) script.
  • Binary Name Matching: Determines if two given names refer to the same individual, outputting a "Yes" or "No" decision.
  • Robustness: Designed to handle common variations in Indian names, including:
    • Spelling variations
    • Abbreviations
    • Token reordering
    • Merged tokens
    • Honorifics (e.g., "-bhai", "-ji")

Performance

Evaluated on a test set of 9581 name pairs, the model demonstrates exceptional performance:

  • Precision: 0.9997
  • Recall: 0.9997
  • F1 Score: 0.9997
  • Accuracy: 0.9998

Usage

The model is easily integrated using the Hugging Face transformers library, with examples provided for direct inference. It is distributed as a merged, full model in bf16 format.

Good For

  • Applications requiring highly accurate name parsing and matching in the context of Indian KYC.
  • Systems needing to reconcile names with variations, abbreviations, and cultural specificities.
  • Developers looking for a compact yet powerful model for specialized name entity resolution tasks.

Finetunes
1 models