A fine-tuned version of Qwen2.5-3B-Instruct optimized for structured tool calling, business analytics, and robust reasoning over noisy real-world tabular data.
Unlike the base model, this version focuses on producing schema-valid JSON tool calls, handling inconsistent CSV structures, and performing multi-step KPI calculations with significantly higher reliability.
Highlights
🧠 Fine-tuned using Supervised Fine-Tuning (SFT) + QLoRA (PEFT)
📊 Optimized for business analytics workflows
🔧 Reliable JSON-RPC tool invocation
📁 Handles noisy and malformed CSV schemas
⚡ 4-bit quantized inference support
🤖 Designed for MCP (Model Context Protocol) tool integration
Model Details
Property
Value
Base Model
Qwen2.5-3B-Instruct
Fine-tuning
SFT + QLoRA
Framework
PyTorch
Libraries
Transformers, PEFT, TRL
Intended Use
Tool Calling, Business Analytics, KPI Reasoning
Quantization
4-bit inference supported
Motivation
Large language models frequently struggle with structured business data because real-world datasets are rarely clean.
Common failure modes include:
malformed JSON tool calls
incorrect function arguments
hallucinated fields
broken numerical reasoning
inability to adapt to inconsistent column names
This project fine-tunes Qwen2.5-3B to improve reliability on these tasks while maintaining low inference latency.
Training
Fine-tuning Method
Supervised Fine-Tuning (SFT)
QLoRA parameter-efficient adaptation
Hugging Face Transformers
PEFT
TRL
PyTorch
Synthetic Data Pipeline
Training data was generated to simulate real-world business datasets by introducing:
inconsistent column names
missing values
malformed schemas
noisy numerical data
adversarial table layouts
ambiguous KPI requests
The objective was to improve robustness against inputs that commonly cause tool-calling failures.
Evaluation
Evaluation was performed on a held-out adversarial benchmark containing noisy business documents and malformed tabular schemas.
Results
Metric
Base Model
Fine-Tuned
Function-calling Schema Compliance
81.2%
99.6%
Reasoning / Output Failures
Baseline
76% Reduction
Inference Latency
—
<3.2 s
Intended Use
This model is designed for applications involving:
Business Intelligence
KPI Calculation
Financial Reporting
Dashboard Generation
CSV Analysis
Structured Tool Calling
Agentic AI Workflows
MCP-based Systems
Tool Calling
The model is trained to generate structured JSON-RPC calls for tools such as:
This model was developed as the reasoning engine for an MCP-powered analytics platform.
It integrates directly with:
services/mcp_service.py
services/tool_service.py
If the fine-tuned weights are unavailable, the application gracefully falls back to an Ollama-hosted model without requiring changes to the surrounding tool pipeline.
