Upadhyee/swar-nlu-gemma3-4b-lora

SWAR-NLU — Gemma 3 4B LoRA Adapters

SWAR-NLU is a multilingual Natural Language Understanding (NLU) system built on Google Gemma 3 4B Instruct and fine-tuned using Low-Rank Adaptation (LoRA) through PEFT.

The project focuses on multilingual intent classification and slot filling across Indian languages and evaluates how training data size and language holdout affect NLU performance.

This Hugging Face repository contains the trained LoRA adapters and the tokenizer artifacts required to load them.

Important: These files are LoRA adapters, not standalone copies of Gemma 3 4B. The original base model is required to use these adapters.

---

Model Details

Property	Value
Base model	google/gemma-3-4b-it
Fine-tuning	LoRA
Framework	Hugging Face Transformers + PEFT
Model type	Causal Language Model
Task	Multilingual NLU
Tasks	Intent classification + slot filling
Languages	Bengali, Hindi, Kannada, Malayalam, Tamil, Telugu
Adapter format	Safetensors
Base model parameters	Gemma 3 4B
Adapter size	~114 MiB per adapter

The adapters were trained using parameter-efficient fine-tuning rather than updating the complete Gemma model.

---

Available Adapters

The repository contains two groups of experiments.

1. Training-Size Experiments

These adapters investigate the effect of increasing the amount of training data.

Adapter	Training Data
r3k	3,000 examples
r9k	9,000 examples
r18k	18,000 examples

r3k

LoRA adapter trained using approximately 3,000 training examples.

r9k

LoRA adapter trained using approximately 9,000 training examples.

r18k

LoRA adapter trained using approximately 18,000 training examples.

---

2. Leave-One-Language-Out Experiments

The leave-one-language-out (LOO) experiments evaluate cross-lingual generalization by holding one language out of training.

Adapter	Held-Out Language
loo_bn-BD	Bengali
loo_hi-IN	Hindi
loo_kn-IN	Kannada
loo_ml-IN	Malayalam
loo_ta-IN	Tamil
loo_te-IN	Telugu

For example, loo_ml-IN represents the experiment where Malayalam is the held-out language.

These adapters are intended to support analysis of how well the model generalizes to a language that was not included in the corresponding training setup.

---

Languages

The project covers the following Indian languages:

Language	Locale
Bengali	bn-BD
Hindi	hi-IN
Kannada	kn-IN
Malayalam	ml-IN
Tamil	ta-IN
Telugu	te-IN

---

Evaluation

The trained adapters are evaluated for multilingual NLU performance.

The evaluation pipeline reports the following metrics:

• Intent Accuracy
• Intent Macro F1
• Slot Micro F1
• Slot Precision
• Slot Recall
• Slot Type F1
• Exact Match
• JSON Validity

The detailed evaluation outputs are maintained separately in the SWAR-NLU source repository.

The Hugging Face repository primarily provides the trained model artifacts required for reproducing inference with the adapters.

---

Repository Structure

swar-nlu-gemma3-4b-lora/
│
├── r3k/
│   ├── adapter_config.json
│   ├── adapter_model.safetensors
│   └── tokenizer artifacts
│
├── r9k/
│   ├── adapter_config.json
│   ├── adapter_model.safetensors
│   └── tokenizer artifacts
│
├── r18k/
│   ├── adapter_config.json
│   ├── adapter_model.safetensors
│   └── tokenizer artifacts
│
├── loo_bn-BD/
│   ├── adapter_config.json
│   ├── adapter_model.safetensors
│   └── tokenizer artifacts
│
├── loo_hi-IN/
│   ├── adapter_config.json
│   ├── adapter_model.safetensors
│   └── tokenizer artifacts
│
├── loo_kn-IN/
│   ├── adapter_config.json
│   ├── adapter_model.safetensors
│   └── tokenizer artifacts
│
├── loo_ml-IN/
│   ├── adapter_config.json
│   ├── adapter_model.safetensors
│   └── tokenizer artifacts
│
├── loo_ta-IN/
│   ├── adapter_config.json
│   ├── adapter_model.safetensors
│   └── tokenizer artifacts
│
└── loo_te-IN/
    ├── adapter_config.json
    ├── adapter_model.safetensors
    └── tokenizer artifacts