Fine-tuning NVIDIA Nemotron-Mini-4B for cricket domain expertise
CricketMind is an end-to-end fine-tuning project that transforms a general-purpose LLM into a cricket laws and match-situation reasoning specialist. It shows that meaningful domain adaptation is possible with ~170 training examples, 35 minutes of GPU time, and under $5.
| Metric | Baseline (Nemotron-Mini-4B) | Fine-Tuned | Improvement |
|---|---|---|---|
| Overall Score | 29.8% | 67.5% | +37.7 pp |
Evaluated on CricketBench v0.1 — a custom benchmark with 20 questions across 4 difficulty tiers, scored using LLM-as-judge methodology.
- Laws Corpus: 8 major MCC cricket laws -> ~160 generated QA pairs
- Match Situations: 50 real-world scenarios processed through response distillation (Claude as teacher model)
- Evaluation Set: CricketBench — 20 questions, 4 difficulty categories, weighted scoring
- Method: QLoRA fine-tuning (rank 16, 4-bit quantization)
- Precision: bfloat16
- Hardware: RunPod A100 (~$4 total cost)
- Time: ~35 minutes
cricketmind/
data/ # Training data and laws corpus
evaluation/ # CricketBench benchmark and scoring
model/ # Model configs and training scripts
scripts/ # Data generation and pipeline utilities
model_card.md # Full model documentation
dataset_card.md # Dataset documentation
- Response Distillation — Using Claude as teacher to generate high-quality training data
- QLoRA — Parameter-efficient fine-tuning with 4-bit quantisation for affordable GPU training
- LLM-as-Judge Evaluation — Automated scoring with confidence thresholds (0.8) and structured rubrics
- Custom Benchmarking — Domain-specific eval with weighted difficulty tiers
git clone https://github.com/brettleehari/cricketmind.git
cd cricketmind
pip install -r requirements.txt
python scripts/generate_qa_pairs.py
python evaluation/run_benchmark.pyFine-tuning requires GPU access (A100 recommended). S
Hariprasad Sudharshan - GitHub - OSS contributor to LlamaIndex
MIT