Llama-3.2-1B-nba-ner-GGUF-Q4_K_M : GGUF

This model is a fine-tuned version of Llama 3.1 1B, optimized for Named Entity Recognition (NER) tasks on NBA sentences. It is specifically designed to extract player and team names from text.

Performances

The model is designed for efficiency: it matches performances of state-of-the-art models 10x larger, while maintaining fast inference times.

Model Comparison

Model	Size (B)	Accuracy (%)	Mean Processing Duration* (s)
Llama-3.2-nba-ner	1	90.1	0.38
Llama3.2	1	32.7	0.79
Gemma2	2	58.3	1.43
Gemma3	4	75.3	0.69
Llama3.1	8	56.5	1.16
Gemma3	12	92.9	2.73

* Benchmarked on an RTX 3060 Laptop GPU using the test part of the pdesj/nba-ner-team-player-1617 dataset.

Dataset

The model was trained using the train part of the pdesj/nba-ner-team-player-1617 dataset. This dataset contains annotated NBA related sentences in english and french, with labeled player and team names, synthetically generated using LLMs, and reviewed by a human annotator.

How to Run

See recommended system prompt for better performance.

Using llama.cpp

./llama.cpp/llama-cli -hf pdesj/Llama-3.2-1B-nba-ner-GGUF-Q4_K_M --jinja

Using ollama

ollama run pdesj/Llama-3.2-1B-nba-ner-GGUF:Q4_K_M

References

• Codebase used to train the model: repo
• Trained with unsloth