Intended use & limitations

Retrieval-augmented Q&A for mining / quarry / blasting professionals. Not a standalone knowledge source — use with the retriever (intfloat/e5-base-v2 + a Qdrant collection). For SDS / safety content, always verify against the cited source PDF; the model is trained to give page numbers for exactly this reason. Domain-specific to Dyno Nobel AU products.

Training

• Base: Qwen/Qwen3-4B
• Method: QLoRA (4-bit nf4), r=16, α=32, dropout=0.05, targets all attn + MLP projections
• Data: 602 synthetic grounded examples — 584 [N]-cited answers (SDS sections, tech specs, guides, case studies) + 18 refusal / safe-decline examples — generated by a teacher model over retrieved context
• Schedule: 3 epochs, lr 1e-4 cosine, full-sequence SFT
• Result: final train_loss 1.48 (2.54 → 1.12), token accuracy 57% → 76%

Files

• *.safetensors — merged fp16 weights (load with transformers)
• dyno-blast-4b-q8_0.gguf — q8_0 GGUF for llama.cpp / Ollama

Prompt format

Grounded system prompt (answer only from numbered SOURCEs, cite [N], refuse if absent) + numbered SOURCE [N] blocks from the retriever, then the question. The exact system prompt and chunk schema are in the companion dataset repo.