kcherry497/dyno-blast-4b
kcherry497/dyno-blast-4b is a 4 billion parameter QLoRA fine-tune of Qwen3-4B, specifically designed for grounded question answering in the blasting and explosives domain. It strictly answers technical and safety questions from retrieved context, attaching citations to every claim, and refusing to answer if information is not present in the provided sources. This model is optimized for retrieval-augmented generation (RAG) systems, particularly for Dyno Nobel Australia's public technical literature.
Loading preview...
Overview
kcherry497/dyno-blast-4b is a specialized 4 billion parameter model, fine-tuned using QLoRA on Qwen3-4B. Its core function is to provide grounded answers to technical and safety questions related to blasting and explosives, drawing exclusively from provided source contexts. A key feature is its ability to cite every claim with [N] + page numbers and to refuse to answer if the information is not found in the given sources.
Key Capabilities
- Grounded Q&A: Answers are strictly derived from provided, numbered
SOURCE [N]blocks. - Citation Generation: Automatically attaches
[N]+ page citations to all generated claims. - Refusal Mechanism: Explicitly trained to refuse questions when the answer is not present in the supplied context.
- Domain-Specific: Optimized for Dyno Nobel Australia's public technical literature, including product data sheets, Safety Data Sheets (SDS), and application guides.
- RAG Component: Designed as the generation half of a Retrieval-Augmented Generation (RAG) system, intended to be paired with a companion vector database like
kcherry497/dyno-blast-4b-rag.
Training Details
The model was trained using QLoRA (4-bit nf4) on 1,674 synthetic grounded examples, including 1,656 cited answers and 18 refusal examples. The training data covers SDS sections, technical specs, application guides, and more, generated by a teacher model over retrieved context. The training involved 3 epochs with a learning rate of 1e-4 cosine, resulting in a final train_loss of 0.97.
Intended Use
This model is intended for retrieval-augmented Q&A for mining, quarry, and blasting professionals. It is not a standalone knowledge source and must be used with a retriever. Users are advised to verify SDS and safety content against the cited source PDF.