HasinduNimesh/qwen3b-finetuned

TEXT GENERATIONPricing:Input $0.32 / Cached $0.064 / Output $1.6Concurrent Unit Cost:1Model Size:3.1BQuant:BF16Context Size:32kTool Calling:SupportedPublished:Mar 10, 2025License:mitArchitecture:Transformer Open Weights Featherless Exclusive Cold

HasinduNimesh/qwen3b-finetuned is a 3.1 billion parameter Qwen2.5-3B-Instruct model, fine-tuned by HasinduNimesh using Unsloth. This model is specifically optimized for multi-hop reasoning, scientific question answering, and retrieval-augmented generation (RAG) using FAISS and BM25 retrieval. It is available in both 4-bit GGUF and 16-bit versions, making it suitable for efficient inference in specialized applications requiring advanced reasoning capabilities.

Loading preview...

Model Overview

This model, HasinduNimesh/qwen3b-finetuned, is a specialized fine-tuned version of the Qwen2.5-3B-Instruct base model. Developed by HasinduNimesh, it leverages Unsloth for efficient fine-tuning and is designed to excel in complex reasoning tasks.

Key Capabilities

  • Multi-hop Reasoning: Optimized using datasets like HotpotQA to handle questions requiring multiple steps of inference.
  • Scientific Q&A: Enhanced for answering scientific questions, potentially through its reasoning and retrieval capabilities.
  • Retrieval-Augmented Generation (RAG): Integrates with FAISS and BM25 retrieval methods, allowing it to generate responses based on retrieved documents.
  • Efficient Deployment: Available in both 4-bit GGUF (for Llama-CPP) and 16-bit versions, offering flexibility for various deployment environments.

Fine-Tuning Details

The model underwent a two-stage fine-tuning process:

  • Supervised Fine-Tuning (SFT): Utilized datasets including HotpotQA and synthetic QA generated from document chunks.
  • Offline Reward Preference Optimization (ORPO): Applied for further improving reasoning performance.

Good For

  • Applications requiring advanced multi-hop reasoning.
  • Systems needing to perform scientific question answering.
  • Integrating with RAG pipelines for enhanced factual grounding.
  • Deployment on resource-constrained devices using its quantized GGUF format.