Elfibolt/gonyai-intent-indicbertv2

Gonyai Konkani Intent Classifier (IndicBERT v2)

This model is a high-accuracy, low-latency intent router for Goan Konkani (kok / gom) in Devanagari script. It is built as the first-stage intent classifier for the Gonyai Konkani AI assistant to route user prompts into specific backend pipelines (Live News APIs, Real-time Weather, RAG Document Context, or Direct LLM generation).

Fine-tuned from ai4bharat/IndicBERTv2-MLM-only.

---

🎯 Target Intent Classes

Label ID	Intent	Description	Example
0	NEWS	Live news, current affairs, sports scores, and political events.	"गोवा विधानसभेंत अधिवेशना वेळार कितें चर्चा जाली?"
1	WEATHER	Real-time weather, temperature, forecasts, rainfall, and sea alerts.	"आयज पणजेंत हवामान कशें आसा?"
2	CONTEXT	Queries referring to an uploaded document, scheme rules, PDFs, or contracts.	"ह्या दस्तऐवजांतले कायदे आनी अटी म्हाका समजावन सांग"
3	DIRECT	General assistant queries, writing assistance, recipes, coding, facts, and conversation.	"माशांच्या कडीची सोपी रेसिपी सांग"

---

📊 Dataset & Training Details

• Dataset Size: 5,300 curated Goan Konkani queries.
• Balanced Hard Negatives (~16% of dataset): Contains boundary cases (e.g. queries mentioning rain/weather words to write poetry or cook food, or mentioning news to request translations) to prevent false-positive routing.
• Dialect Curation: Cleaned and audited specifically to eliminate Marathi language contamination (e.g. Marathi -चा vs Konkani -चो, Marathi कसा/कसे vs Konkani कसो/कशें, Marathi झाला/झाली vs Konkani जालो/जाली).
• Base Architecture: IndicBERT v2 (110M parameters, native Devanagari vocabulary).
• Hardware: Fine-tuned on Kaggle GPU with mixed precision (fp16).

---

🚀 How to Use

1. Using Hugging Face pipeline

from transformers import pipeline

classifier = pipeline(
    "text-classification",
    model="your-username/gonyai-intent-indicbertv2"
)

query = "आयज पणजेंत कितलें तापमान आसा?"
result = classifier(query)
print(result)
# Output: [{'label': 'WEATHER', 'score': 0.998...}]