ekacare/parrotlet-v-2.5-pro

Model Overview

ekacare/parrotlet-v-2.5-pro is a specialized multimodal model developed by Eka Care, fine-tuned from Google's gemma-4-E2B-it (Gemma 4, E2B) with approximately 2 billion effective parameters and a 32768 token context length. It is designed for structured, schema-driven extraction from Indian medical records, including prescriptions, invoices, lab reports, and discharge summaries, as well as PII detection. The model's language backbone was fully fine-tuned, while vision and audio components remained frozen.

Key Capabilities

• Schema-driven Data Extraction: Extracts structured data using specific tags like <prescription>, <invoice>, <lab_report>, and <discharge_summary>.
• PII Detection: Capable of identifying and extracting Personally Identifiable Information (PII) across various document types.
• High Performance: Achieves an overall rubric score of 85.72, outperforming Gemini 3.6 Flash, GPT-5.6 Terra, Claude Sonnet 5, and GPT-5.6 Luna on internal benchmarks for medical document intelligence. It shows a significant improvement of ~40 points over its base model.
• Strong Grounding: Excels in grounding, ensuring extracted information is well-supported by the source document.

Good For

• Developers needing to automate data extraction from diverse Indian medical documents.
• Applications requiring high accuracy in PII detection within medical contexts.
• Use cases where structured output from medical records is critical, such as healthcare administration or research.

Limitations

• Weakest on invoices compared to frontier models, particularly with handwritten bills.
• Performance on complex/non-tabular lab report layouts (e.g., echo, ECG) shows remaining headroom.
• Evaluated only on an internal dataset; not benchmarked on out-of-distribution layouts, languages, or non-Indian medical record formats.
• Not a diagnostic or clinical decision-making tool; human review of outputs is essential for compliance-sensitive workflows.