YouTube entry: Rosettia video

rosettia-quy-gspo-nllb13b-merged (Spanish → Chanka/Ayacucho Quechua)

Standalone merged NLLB-200-1.3B fine-tuned with GSPO reinforcement learning for spa→quy. This is facebook/nllb-200-1.3B with the GSPO LoRA merged in — load directly, no PEFT needed.

ChrF 45.53 (w0), standalone, AmericasNLP 2021 spa→quy test. Full results, figures, scorecard, methodology, and honest caveats (which number is pre-registered vs best-found) are in the adapter repo: https://huggingface.co/Thermostatic/rosettia-quy-gspo-nllb13b-lora

Usage

import torch
from transformers import AutoTokenizer, AutoModelForSeq2SeqLM
tok = AutoTokenizer.from_pretrained("Thermostatic/rosettia-quy-gspo-nllb13b-merged", src_lang="spa_Latn", tgt_lang="quy_Latn")
m = AutoModelForSeq2SeqLM.from_pretrained("Thermostatic/rosettia-quy-gspo-nllb13b-merged", torch_dtype=torch.bfloat16).cuda().eval()
bos = tok.convert_tokens_to_ids("quy_Latn")
enc = tok("No se por que sucedio eso.", return_tensors="pt").to("cuda")
print(tok.batch_decode(m.generate(**enc, forced_bos_token_id=bos, num_beams=5, max_new_tokens=128), skip_special_tokens=True)[0])

Research-grade; single-benchmark single-reference ChrF; no human evaluation. See the adapter repo for limitations.

Authors & contributions

A two-person SomosNLP hackathon project:

• Estefanía Espinosa Fernández — data curation, and the initial Qwen3.5 LoRA experiments (comparing DoRA, rsLoRA and LoRA, and exploring data mixes).
• Irving Ernesto Quezada Ramírez (irvingernesto.com) — the subsequent modeling through the final system: synthetic-data distillation, the NLLB pipeline, GSPO reinforcement learning, decoding/ensembling, evaluation, and release.

The project was a close collaboration; both contributions were essential to the result.

Links

• Code: https://github.com/Sekinal/rosettia-chanka
• NLLB/M2M-100 support for vLLM (our fork): https://github.com/Sekinal/vllm/tree/add-nllb-m2m100-support
• Data: https://huggingface.co/datasets/Thermostatic/rosettia-chanka-data

Compute & funding

All compute was self-funded out of pocket by Irving Ernesto Quezada Ramírez. The project used roughly ~90 hours of L40S and ~90 hours of A100 GPU time in total, the overwhelming majority of it on experimentation. Training the important models took about ~8 hours of A100 time, plus roughly ~4 hours of A100 for vLLM inference. All told, the whole project cost about ~$100.