ratnasekhar/earnings-copilot-phi3-merged
The ratnasekhar/earnings-copilot-phi3-merged model is a 4 billion parameter Phi-3.5-mini-instruct variant, fine-tuned using QLoRA for structured KPI extraction from SEC filings. It excels at extracting financial metrics like Revenue and EPS into JSON format, providing source quotes, and explicitly indicating when data is unverified. This model is specifically designed to prevent hallucination by returning 'UNVERIFIABLE' confidence when information is not present in the provided text chunks.
Loading preview...
Overview
This model, ratnasekhar/earnings-copilot-phi3-merged, is a 4 billion parameter version of Microsoft's Phi-3.5-mini-instruct, fine-tuned using QLoRA. It is a key component of the Earnings Intelligence Copilot, a multi-agent system designed to process SEC filings and generate citation-grounded investment memos. The model's primary function is to extract structured financial KPIs from text chunks.
Key Capabilities
- Structured KPI Extraction: Extracts financial metrics such as Revenue, Gross Margin, Operating Income, EPS, and Free Cash Flow into a JSON format.
- Hallucination Prevention: Explicitly returns
{"confidence": "UNVERIFIABLE"}when requested data is not found in the input text, rather than inventing information. - Source Grounding: Every extracted answer includes a
source_quotefield, directly linking the output to the original filing text for verifiability. - Balanced Training: Fine-tuned on a balanced dataset (50% verifiable, 50% unverifiable examples) to ensure robust performance in both data presence and absence scenarios.
Use Cases
This model is ideal for applications requiring precise and verifiable extraction of financial data from regulatory documents. It is particularly well-suited for:
- Automating the extraction of key performance indicators from SEC 10-K and 10-Q filings.
- Building financial analysis tools that require structured data input.
- Developing systems where preventing hallucination and ensuring data traceability are critical.