sascha-frank-ai-research/tsft-rag-qwen2.5-14b-instruct
The sascha-frank-ai-research/tsft-rag-qwen2.5-14b-instruct is a 14 billion parameter Qwen2.5-based model, developed by Sascha Frank, specifically fine-tuned for Retrieval-Augmented Generation (RAG) tasks. It excels in answering from supplied context, abstaining when context is insufficient, and generating structured JSON and citation-oriented responses. This model is optimized for research and experimental RAG workflows, particularly for context-grounded question answering in German.
Loading preview...
TSFT-RAG Qwen2.5-14B-Instruct Overview
This model is a 14 billion parameter, full-parameter fine-tuned derivative of Qwen/Qwen2.5-14B-Instruct, developed by Sascha Frank as part of the TSFT-RAG model family. It is specifically designed to enhance Retrieval-Augmented Generation (RAG) systems, focusing on improving several key behaviors critical for practical RAG applications. The model supports a context length of 32,768 tokens, though it was trained with a sequence length of 1,024 tokens, primarily using a German-language corpus.
Key Capabilities and Improvements
The fine-tuning process targeted specific RAG challenges, resulting in notable improvements:
- Context-grounded answering: Enhanced ability to answer questions based solely on provided context.
- Abstention behavior: Improved capacity to abstain from answering when the context does not support a response, with significant gains in hard-negative handling (correctly rejecting 250 of 708 cases compared to 21 for the base model).
- Structured output: Better adherence to task-specific output requirements, including valid JSON generation.
- Citation-oriented responses: Strong improvements in generating source-aware and citation-related answers, with citation precision and recall increasing from 0.0000 to 0.2178.
While showing substantial gains in these areas, the model exhibits small declines in grounded QA and some analysis metrics, indicating task-specific trade-offs. The training involved full-parameter supervised fine-tuning over 3 epochs with a learning rate of 1e-6.
Intended Use Cases
This model is primarily intended for research and experimental use in RAG-oriented workflows. Suitable applications include:
- Context-grounded question answering.
- Evaluation of abstention behavior in RAG systems.
- Generation of structured RAG responses.
- Source-aware answer generation and comparison of base vs. fine-tuned models.
It is crucial to use this model with retrieved evidence or explicitly supplied context, as it is not designed to be a standalone factual knowledge source. The model's primary training language is German, which may influence its behavior in other languages.