thadou-kuki-qwen3-4b-lora

Thadou-Kuki ↔ English translator (Qwen3-4B LoRA, v1)

First open translation model for Thadou-Kuki (Thado Chin, ISO 639-3 tcz), spoken in Manipur, Assam, Nagaland and Chin State.

• adapter_model.safetensors: LoRA (r=32) on unsloth/Qwen3-4B-Instruct-2507
• gguf/Qwen3-4B-Instruct-2507.Q4_K_M.gguf: merged, 4-bit, for Ollama / llama.cpp (2.5 GB)
• gguf/Modelfile: Ollama config

Usage

Prompt formats seen in training:

English to Thadou-Kuki:

God loves you.

Thadou-Kuki to English:

Kapa le kanu chu inn ah aum uve.

Ollama: download gguf/, then ollama create thadou -f Modelfile.

Transformers + PEFT:

from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel
repo = "Prompt48/thadou-kuki-qwen3-4b-lora"
tok = AutoTokenizer.from_pretrained(repo)
base = AutoModelForCausalLM.from_pretrained("unsloth/Qwen3-4B-Instruct-2507", torch_dtype="auto", device_map="auto")
model = PeftModel.from_pretrained(base, repo)

System prompt used in training: You are an expert translator between Thadou-Kuki (Thado Chin) and English.

Training

• Data: 30,814 Bible verses aligned with the World English Bible and KJV. 95,241 examples covering both directions.
  • Thadou text: Pathen Lekhabu Theng (Chongthu dialect, © 2020 Chongthu Bible Translation Team, CC BY-SA 4.0, via eBible.org).
  • Plus 13k verses from an older Thadou Bible (gospelgo.com).
• 16-bit LoRA r=32, 1 epoch, lr 2e-4, batch 16, Unsloth, loss on responses only. Trained on an A40 in 1h48m.

Evaluation

Books held out of training: Ruth, Jonah, Philippians, Jude, 3 John. 200 sentences.

Limitations

• Only Bible text went into training, so everyday sentences drift into scripture wording.
• Mostly Chongthu-dialect spelling.
• Not checked by native speakers. Don't use it for anything important without human review.