ergar/monolingual-reinforcement-learning-mt

Reference-Free RL Fine-Tuning for MT — LoRA Adapters

LoRA adapters (rank 16, q_proj/v_proj) from "Reference-Free Reinforcement Learning Fine-Tuning for MT: A Seq2Seq Perspective", applying GRPO to NLLB-200 (600M/1.3B) with a reference-free reward (LaBSE + COMET-Kiwi). Code: https://github.com/erngarcia/reference-free-ft-rl-mt

Every folder here corresponds to one training run's best intermediate checkpoint (highest intermediate chrF++ on a held-out dev subset), renamed from the code's own adapters_best/ to adapters/ for clarity. The directory structure mirrors the project's own results/ layout — see the code repo's README for what each run tag (exp_a_*, exp_b_*, cascade_*, ccnews_scaling_*, data_size_ablation_*, etc.) means.

Loading an adapter

from peft import PeftModel
from transformers import AutoModelForSeq2SeqLM

base = AutoModelForSeq2SeqLM.from_pretrained("facebook/nllb-200-distilled-600M")
adapter = PeftModel.from_pretrained(
    base,
    "ergar/monolingual-reinforcement-learning-mt",
    subfolder="results/exp_a_600m_en_arb_Arab/adapters/arb_Arab",
)

Swap the subfolder path and base model (600M vs. 1.3B — check the run tag) for any other language/condition in the repo.

Scope

This repo excludes results/legacy_original_submission/ adapters (superseded by this project's rerun where applicable — see the code repo's own note on that directory).