🦊 klusai/tf2-12b — LoRA-tuned EN→RO Translator (Fables) — GGUF build

klusai/tf2-12b is a Gemma 3 12B (Instruct) based translator, fine-tuned with LoRA on 15,000 English→Romanian fable pairs (synthetic, generated with GPT-o3). This repository distributes the merged weights in GGUF format for fast inference with llama.cpp-compatible runtimes.

Focus: faithful, fluent EN→RO translation of short moral fables, preserving tone, structure, and morals (with proper Romanian diacritics).

✨ Highlights

Task: EN→RO translation for narrative prose (fables).
Training: Parameter-efficient LoRA; adapters merged into base and exported to GGUF.
Data: 15k synthetic pairs generated by GPT-o3.
Format: GGUF (llama.cpp). Multiple quantizations may be provided (e.g. Q5_K_M, F16).

🧱 I/O Contract

Input: English fable paragraph(s) + concise instruction to translate to Romanian.
Output: Romanian translation that is faithful, fluent, and keeps the moral explicit.

Prompting tip: “Translate to Romanian with proper diacritics. Preserve meaning, tone, and moral. Avoid adding content.”

🧪 Training Summary

Base model: Gemma 3 12B (Instruct)
Method: LoRA fine-tuning; adapters merged for deployment
Dataset: 15k EN→RO fable translations (GPT-o3 generated)

✅ Intended Use

Education/NLP demos, EN→RO machine translation of narrative content, style-preserving localization.

⚠️ Limitations

Tuned on fables; may underperform on legal/medical/colloquial domains.
Synthetic data can carry stylistic biases or occasional literalism.
Always add human review for production use.

🔑 Licensing

This GGUF build: MIT
Use must also comply with the Gemma base model license/terms.

📝 Changelog

v1.0 (GGUF) — Initial public GGUF release (LoRA-merged). Multiple quantizations provided (e.g., Q5_K_M).

klusai
/

tf2-12b-gguf