Koyna-V2-1b-instruct

A fine-tuned version of gemma-3-1b-it trained on OCR-scanned BSCE Agriculture textbooks. It is a bilingual model fluent in Marathi and English, designed for agriculture domain tasks, including answering syllabus-based questions, general real-world farming conversations.

---

🧠 Model Details

📝 Description

• Model Name: Koyna-V2-1b-instruct
• Base Model: google/gemma-3-1b-it
• Architecture: Gemma 3B Instruction-tuned
• Fine-tuned by: Govind Barbade
• Languages: Marathi (mr), English (en)
• License: apache-2.0
• Use Case: Conversational, QA, and instruction-following for farming/agriculture education

---

📦 Model Sources

• 🤗 Original repo: Govind222/Koyna-V2-1b-instruct

---

💬 Uses

✅ Direct Use

• Answering questions from BSCE Agriculture syllabus
• Conversational agent in Marathi + English
• Educational assistant for rural/agri tech

🚫 Out-of-Scope Use

• Medical, legal, or critical decision-making
• Bias-free or politically sensitive generation without supervision

---

⚠️ Bias, Risks, and Limitations

• Trained on scanned OCR text; may contain noise or formatting errors
• May reflect biases present in the original academic materials
• Not tested on adversarial queries

🔎 Recommendations

Use in supervised educational or non-critical contexts. Validate outputs before use in production/agricultural planning.

---

🚀 Getting Started

from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained("Govind222/Koyna-V2-1b-instruct")
tokenizer = AutoTokenizer.from_pretrained("Govind222/Koyna-V2-1b-instruct")

inputs = tokenizer("माझ्या ऊस पिकासाठी कोणते खत उपयुक्त आहे?", return_tensors="pt")
outputs = model.generate(**inputs)
print(tokenizer.decode(outputs[0]))

🧪 Training Details

📚 Dataset

• Manually collected and OCR-scanned BSCE Agriculture textbooks (Marathi)
• Chapters include: Agronomy, Soil Science, Horticulture, Entomology, Plant Pathology

📊 Evaluation

Model was evaluated qualitatively on syllabus-based QA and conversational prompts. Further benchmarking in progress.

🧰 Technical Specs

• Model Type: Causal LM (Decoder only)
• Base Architecture: Gemma 1B Instruction-tuned
• Quantized Versions: GGUF available in Q2_K, Q4_K_M, Q8_0, F16, etc.

🙏 Acknowledgements

Thanks to: Google for the base model My team and resources at home for enabling this project

📫 Contact

Author: Govind Barbade Email: [email protected] Hugging Face Profile: https://huggingface.co/Govind222