sascha-frank-ai-research/tsft-rag-qwen2.5-1.5b-instruct

TEXT GENERATIONPricing:Input $0.04 / Cached $0.008 / Output $0.08Concurrent Unit Cost:1Model Size:1.5BQuant:BF16Context Size:32kTool Calling:SupportedPublished:Jul 20, 2026License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Cold

TSFT-RAG Qwen2.5-1.5B-Instruct is a 1.5 billion parameter causal language model developed by sascha-frank-ai-research, derived from Qwen/Qwen2.5-1.5B-Instruct. This model is specifically fine-tuned for Retrieval-Augmented Generation (RAG) tasks, focusing on improving context-grounded answering, abstention behavior, and structured output validity. It supports a context length of 32,768 tokens and is primarily trained on a German-language corpus. The model excels in hard-negative handling and generating citation-aware responses, making it suitable for research and experimental RAG workflows.

Loading preview...

TSFT-RAG Qwen2.5-1.5B-Instruct Overview

This model is a 1.5 billion parameter variant from the TSFT-RAG (Task-Specific Full Fine-Tuning for Retrieval-Augmented Generation) series by sascha-frank-ai-research. It is a full-parameter fine-tuned derivative of Qwen/Qwen2.5-1.5B-Instruct, designed to investigate the impact of supervised fine-tuning on RAG system behavior. The model supports a substantial context length of 32,768 tokens and was trained on a German-language corpus.

Key Capabilities & Differentiators

  • Enhanced RAG Performance: Specifically fine-tuned to improve several critical RAG behaviors.
  • Improved Hard-Negative Handling: Demonstrates significant gains in correctly abstaining when context does not support an answer, with a +0.3362 difference in score compared to the base model.
  • Structured Output & Citation: Shows substantial increases in JSON validity (+0.1386) and citation-related metrics (precision and recall both +0.6502).
  • Full Fine-Tuning: Utilizes full-parameter supervised fine-tuning, rather than adapter-based methods, for comprehensive behavioral modification.
  • Research-Oriented: Part of a broader research project comparing base and fine-tuned configurations across different model families and scales.

Good for

  • Context-grounded question answering.
  • Evaluating abstention behavior in RAG systems.
  • Generating structured RAG responses (e.g., JSON).
  • Source-aware and citation-oriented answer generation.
  • Comparative studies between base and fully fine-tuned small language models.
  • Local or resource-constrained RAG experiments, particularly with German content.

Finetunes
1 models