GiorgiGE/Kolkha-Mini-Georgian

Kolkha-Mini

Kolkha-Mini is a lightweight language model fine-tuned to specialize in the Georgian language.
It is intended as an early-stage foundation model for Georgian-focused NLP work.

This model prioritizes coherence and language exposure over grammatical perfection and should be treated as a base to build upon, not a production-ready assistant.

---

Base Model

• Qwen/Qwen3-1.7B

---

Fine-Tuning Overview

• Method: QLoRA (4-bit)
• Training type: Causal Language Modeling
• Epochs: 2
• Context length: 1024 tokens
• Optimizer: paged AdamW (8-bit)
• Scheduler: cosine
• Precision: FP16 compute, NF4 quantized base during training

The final model provided here is a fully merged FP16 model (no LoRA adapters required).

---

Training Details (High-Level)

• Base model loaded in 4-bit NF4 using bitsandbytes
• LoRA applied to all major attention and MLP projection layers:
  • q_proj, k_proj, v_proj, o_proj
  • gate_proj, up_proj, down_proj
• Dataset manually packed into fixed 1024-token blocks to maximize GPU utilization
• Chat templates applied prior to tokenization
• Gradient checkpointing enabled for stability

Training was intentionally kept simple and stable, favoring correctness over experimental tricks.

---

Current Capabilities & Limitations

What it does well

• Produces coherent Georgian text
• Understands Georgian sentence structure
• Serves as a solid starting point for further fine-tuning

Known issues

• Grammatically incorrect sentences are common
• Occasional hallucinations
• Sometimes invents non-existent words
• Not instruction-tuned or safety-aligned

These issues are expected given dataset size and training duration.
Performance is expected to improve significantly with a larger and cleaner dataset.

---

Intended Use

• Georgian language research
• Further fine-tuning
• Dataset experimentation
• Low-resource language modeling

Not recommended for:

• Production deployment
• High-stakes or factual tasks
• Safety-critical applications

---

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained(
    "GiorgiGE/Kolkha-Mini-Georgian",
    torch_dtype="auto"
)

tokenizer = AutoTokenizer.from_pretrained(
    "GiorgiGE/Kolkha-Mini-Georgian"
)