pankajpandey-dev/gemma-3-1b-hindi-instruct-lora

🇮🇳 Gemma-3-1B Hindi Instruct — LoRA Adapter

LoRA adapter (r=32) for google/gemma-3-1b-it — the method artifact behind pankajpandey-dev/gemma-3-1b-hindi-instruct. To run, load the base model and apply this adapter; for direct use prefer the merged model or GGUF.

Apply

from peft import PeftModel
from transformers import AutoModelForCausalLM
base = AutoModelForCausalLM.from_pretrained("google/gemma-3-1b-it")
model = PeftModel.from_pretrained(base, "pankajpandey-dev/gemma-3-1b-hindi-instruct-lora")

Part of my 🇮🇳 Hindi LLM Series — weekly experiments adapting small models to Indian languages.

Available formats

Repo	Format	Use
...-hindi-instruct	Merged 16-bit	Transformers
...-hindi-instruct-GGUF	Q4_K_M / Q5_K_M / Q8_0	Ollama, llama.cpp, CPU
...-hindi-instruct-lora	LoRA adapter	Method artifact

Training

• Base: google/gemma-3-1b-it (text-only path)
• Method: LoRA (r=32, α=32, all attn+MLP projections), response-only loss
• Data: AI4Bharat indic-instruct-data-v0.1 — anudesh + dolly (Hindi), chrF≥50 filtered, balanced 6k
• Schedule: 2 epochs, LR 2e-4, effective batch 8 · single T4 (Kaggle, free), fp32, ~167 min · Unsloth + TRL

Recommended decoding: temperature=0.4, top_p=0.9, repetition_penalty=1.3.

Evaluation

प्रश्न: एक छोटे बच्चे को गुरुत्वाकर्षण सरल हिंदी में समझाइए। उत्तर: PASTE_YOUR_BEST_CLEAN_OUTPUT_HERE

Limitations

A 1B model — Hindi fluency is solid; coherence/factual reliability are bounded by scale. Best for short instructions, simple Q&A, and edge/demo use. A Gemma-3-4B Hindi version is the planned next step.

Credits

Base model © Google, used under the Gemma license. Data: AI4Bharat. Fine-tuning: Unsloth.