WMT26Anon/hymt2-7b-sft

Hy-MT2-7B — SFT LoRA adapter

Supervised fine-tuning (SFT) LoRA adapter over tencent/Hy-MT2-7B for English–Russian specialized-terminology translation. Anonymous submission to the WMT26 research track.

Training

• Data: 10,000 (source, chosen) pairs mined from WikiMatrix and ParaCrawl, filtered for terminology headroom (see paper §Resources)
• Objective: SFT on chosen translations with the retrieval-augmented glossary prompt (same prompt as at inference)
• LoRA rank 64, α = 128, dropout 0.05
• 2 epochs, learning rate 1e-4, greedy decoding at inference
• Target modules: q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj

Usage

from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer

base = AutoModelForCausalLM.from_pretrained('tencent/Hy-MT2-7B',
                                            trust_remote_code=True)
model = PeftModel.from_pretrained(base, 'WMT26Anon/hymt2-7b-sft')
tok = AutoTokenizer.from_pretrained('tencent/Hy-MT2-7B',
                                    trust_remote_code=True)

Code

Inference pipeline, KB, test sets, and evaluation scripts: https://anonymous.4open.science/r/RAG_System_for_Specialized_Terms-18BB/