OscarFeMa/ExpertiaChemistry-r16

ExpertiaChemistry (Q4_K_M)

Spanish-first chemistry specialist: formal definitions + formulae where applicable. Fine-tuned from microsoft/Phi-4-mini-reasoning (MIT) with QLoRA r16 on 45k pairs (Wikidata definitional bulk + SPARQL P274 formulae + PubChem compounds + StackExchange accepted answers + Wikipedia leads), 3 epochs on RTX 3070 (seq1024, BF16).

Evaluation (500 held-out samples)

	Perplexity
Base	823.3
+ adapter	60.1 (-92.7%)

Production: 0.87 quality, 100% cycle success in the Expertia pipeline.

Variants in this repo

File	Size	Needs	Use
expertia-chemistry-q4_k_m.gguf	2.5GB	6GB VRAM	Daily inference (Ollama)
expertia-chemistry-f16.gguf	7.7GB	16GB VRAM	Max quality inference
fp16/	7.2GB	—	Base for further fine-tuning

Usage (Ollama)

ollama create expertia-chemistry -f Modelfile-ExpertiaChemistry-Q4F

Note: GGUF uses gpt-2 pre-tokenizer; set num_ctx 8192. Create with lowercase name and real (non-junction) FROM path on Ollama ≥0.33.3.

Limitations

Narrow domain specialist (chemistry). Adapter weights: CC-BY-NC-4.0. Base: MIT.