sascha-frank-ai-research/tsft-rag-qwen2.5-14b-instruct

TEXT GENERATIONPricing:Input $0.431 / Cached $0.0862 / Output $1.12Concurrent Unit Cost:1Model Size:14.8BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Jul 20, 2026License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Cold

The sascha-frank-ai-research/tsft-rag-qwen2.5-14b-instruct is a 14.8 billion parameter Qwen2.5-based model, fully fine-tuned by Sascha Frank for Retrieval-Augmented Generation (RAG) systems. It specializes in improving RAG behaviors such as answering from supplied context, abstaining when context is insufficient, and generating structured JSON or citation-oriented responses. This model is particularly optimized for hard-negative handling and citation-related accuracy within RAG workflows, primarily trained on a German-language corpus.

Loading preview...

TSFT-RAG Qwen2.5-14B-Instruct: RAG-Optimized Fine-Tune

This model is a 14.8 billion parameter, full-parameter fine-tuned derivative of Qwen/Qwen2.5-14B-Instruct, developed by Sascha Frank. It is part of the TSFT-RAG model family, specifically designed to investigate and enhance Retrieval-Augmented Generation (RAG) system behaviors through supervised fine-tuning.

Key Capabilities and Improvements

The fine-tuning process focused on several critical aspects for RAG systems, showing significant gains in specific areas:

  • Context-Grounded Answering: Improved ability to answer questions based only on provided context.
  • Abstention Behavior: Enhanced capacity to abstain from answering when the context does not support a response, with a substantial improvement in hard-negative handling (correctly rejecting 250 of 708 cases compared to 21 of 708 for the base model).
  • Structured Output: Better adherence to task-specific output requirements, including valid structured JSON output.
  • Citation-Oriented Responses: Improved generation of source-aware and citation-related responses, with notable gains in citation precision and recall.

While excelling in these areas, the model shows minor trade-offs in grounded QA and some analysis metrics. The training was conducted on a task-specific German-language corpus.

Intended Use Cases

This model is primarily intended for research and experimental use in RAG-oriented workflows, particularly for:

  • Context-grounded question answering.
  • Evaluating abstention behavior and structured RAG response generation.
  • Source-aware answer generation and comparison of base vs. fine-tuned LLMs in RAG.

It should always be used with retrieved evidence or explicitly supplied context and is not designed as a standalone factual knowledge source. The model inherits limitations from its base model and training data, and its primary language of training is German.