ExpertiaMath (Q4_K_M)
Spanish-first mathematics specialist: formal definitions + formulae, no web opinion.
Fine-tuned from microsoft/Phi-4-mini-reasoning (MIT) with QLoRA r16 on 45k
Wikidata-derived definition/formula pairs (+5k validation), 3 epochs on RTX 3070.
Evaluation (500 held-out samples)
| Perplexity | |
|---|---|
| Base | 827.1 |
| + adapter | 63.8 (-92.3%) |
Production: 0.89 quality, 100% cycle success, ~310 pkgs/cycle in the Expertia pipeline.
Variants in this repo
| File | Size | Needs | Use |
|---|---|---|---|
expertia-math-q4_k_m.gguf |
2.5GB | 6GB VRAM | Daily inference (Ollama) |
expertia-math-f16.gguf |
7.7GB | 16GB VRAM | Max quality inference |
fp16/ |
7.2GB | — | Base for further fine-tuning |
Usage (Ollama)
ollama create ExpertiaMath -f Modelfile-ExpertiaMath-Q4
Note: GGUF uses gpt-2 pre-tokenizer (Ollama ≤0.33.3 does not know phi-3);
set num_ctx 8192 (Qwen-family 256K default loads 43GB otherwise).
Limitations
Narrow domain specialist (mathematics). Not a general assistant. Adapter weights: CC-BY-NC-4.0. Base model: MIT (Microsoft).
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support
Model tree for OscarFeMa/ExpertiaMath-r16
Base model
microsoft/Phi-4-mini-reasoning