sascha-frank-ai-research/tsft-rag-qwen2.5-7b-instruct

TEXT GENERATIONPricing:Input $0.4 / Cached $0.08 / Output $0.8Concurrent Unit Cost:1Model Size:7.6BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Jul 20, 2026License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Cold

TSFT-RAG Qwen2.5-7B-Instruct is a 7 billion parameter model developed by Sascha Frank, derived from Qwen/Qwen2.5-7B-Instruct, and fine-tuned for Retrieval-Augmented Generation (RAG) tasks. This model, part of the TSFT-RAG family, focuses on improving RAG system behaviors such as answering from supplied context, abstention when context is insufficient, structured JSON output, and source-aware responses. It supports a context length of 32,768 tokens and was primarily trained on a German-language corpus, showing significant gains in hard-negative handling and JSON validity.

Loading preview...

TSFT-RAG Qwen2.5-7B-Instruct Overview

This model is a 7 billion parameter, full-parameter fine-tuned derivative of Qwen/Qwen2.5-7B-Instruct, developed by Sascha Frank. It is part of the TSFT-RAG (Task-Specific Full Fine-Tuning for Retrieval-Augmented Generation) model family, which aims to investigate how supervised fine-tuning impacts RAG system behavior across various architectures.

Key Capabilities and Differentiators

The model was specifically trained to enhance several critical aspects of RAG systems:

  • Context-grounded answering: Providing answers strictly based on the supplied context.
  • Abstention: Declining to answer when the context does not support a response. The model shows significant improvement in hard-negative handling, correctly rejecting 391 out of 708 cases compared to the base model's 56.
  • Structured output: Generating valid JSON output, with JSON validity improving from 0.5739 to 0.9069.
  • Source-aware and citation-oriented responses: Producing answers that acknowledge sources, with citation precision and recall improving from 0.0000 to 0.8300.
  • Task-specific instruction following: Adhering to specific output requirements.

While showing strong gains in these areas, some metrics like grounded QA and certain analysis tasks experienced slight declines. The model was trained using full-parameter supervised fine-tuning on a German-language corpus, supporting a context length of 32,768 tokens.

Intended Use Cases

This model is designed for research and experimental use in RAG-oriented workflows. Suitable applications include:

  • Context-grounded question answering.
  • Evaluating abstention behavior in RAG systems.
  • Generating structured RAG responses.
  • Comparing base and fine-tuned language models for RAG.
  • Local or server-based RAG experiments.

It is crucial to use this model with explicitly supplied context, as it is not intended as a standalone factual knowledge source. The primary training language is German, which may influence performance in other languages.

Finetunes
1 models