Modelpublic

JuhaD2/sports-science-3b-gguf

sourceHugging Facellama3.2updated 7mo agoView on Hugging Face

2likes43downloads

Model Card

Sports Science AI — Llama 3.2 3B Instruct (Q4KM GGUF)

A fine-tuned version of meta-llama/Llama-3.2-3B-Instruct specialised in sports science, trained to answer questions about HRV analysis, training load management, recovery science, biomechanics, and sports nutrition.

The model is quantized to Q4KM GGUF format for efficient on-device inference via llama.cpp and llama.rn (React Native).

Model Details

Property	Value
Base model	meta-llama/Llama-3.2-3B-Instruct
Fine-tuning method	QLoRA (PEFT)
LoRA rank (r)	16
LoRA alpha	32
Target modules	qproj, kproj, vproj, oproj, gateproj, upproj, down_proj
Training epochs	2
Training pairs	22,068
Validation pairs	2,453
Learning rate	2e-4 (cosine schedule)
Quantization	Q4KM (llama.cpp)
File size	~1.9 GB

Files

File	Size	Description
`model_q4km.gguf`	~1.9 GB	Q4KM quantized model — use this for inference
`tokenizer.json`	~16 MB	Tokenizer vocabulary
`tokenizer_config.json`	small	Tokenizer settings
`special_tokens_map.json`	small	Special token definitions

Training Dataset

The model was fine-tuned on a custom instruction-response dataset derived from sports science research papers across 5 domains:

Domain	Pairs	%
HRV Analysis	~7,700	35%
Training Load	~5,100	23%
Recovery Science	~3,900	18%
Biomechanics	~2,600	12%
Nutrition & Performance	~2,600	12%

Source papers drawn from: International Journal of Sports Physiology and Performance, British Journal of Sports Medicine, Journal of Sports Sciences, Medicine & Science in Sports & Exercise, Sports Medicine.

Usage

llama.cpp (desktop/server)

bash

./llama-cli \
    -m model_q4km.gguf \
    --gpu-layers 99 \
    -n 300 \
    -p "<|begin_of_text|><|start_header_id|>user<|end_header_id|>What is VO2max and how is it measured?<|eot_id|><|start_header_id|>assistant<|end_header_id|>"

LM Studio

Load model_q4km.gguf directly from the Local Models tab.

llama.rn (React Native)

typescript

import { initLlama } from 'llama.rn';
import RNFS from 'react-native-fs';

const modelPath = `${RNFS.DocumentDirectoryPath}/model_q4km.gguf`;

// Download on first launch
await RNFS.downloadFile({
  fromUrl: 'https://huggingface.co/JuhaD2/sports-science-3b-gguf/resolve/main/model_q4km.gguf',
  toFile: modelPath,
}).promise;

// Load model
const context = await initLlama({
  model: modelPath,
  n_gpu_layers: 99,
});

// Run inference
const result = await context.completion({
  messages: [{ role: 'user', content: 'How does training load affect recovery?' }],
  n_predict: 300,
});

console.log(result.text);

transformers (Python)

The adapter weights are not included here. For Python use, load the base model and apply the original LoRA adapter separately.

Intended Use

—Sports science question answering
—Athlete monitoring and load management guidance
—HRV interpretation assistance
—Recovery protocol recommendations
—Educational tool for coaches and sports scientists

Out of Scope

—Medical diagnosis or treatment
—General-purpose question answering outside sports science
—Real-time physiological monitoring

Limitations

—Responses are based on training data from academic papers up to early 2025
—Should not replace consultation with qualified sports medicine professionals
—Performance on highly specialised sub-topics may vary

License

Built on Llama 3.2 — subject to the Llama 3.2 Community License.