rosettia-quy-v30b-9b-merged

YouTube entry: Rosettia video

rosettia-quy-v30b-9b-merged

Spanish → Chanka / Ayacucho Quechua (quy) translation model. A merged (LoRA → full-weights) fine-tune of Qwen3.5-9B, and the strongest single model in the Rosettia Chanka project. On the AmericasNLP 2021 spa→quy test it sets a new state of the art for the benchmark.

Results — AmericasNLP 2021 spa→quy test (1003 lines, ChrF, sacrebleu word_order=0)

Note: many shared-task papers report ChrF++ (word_order=2), which is not directly comparable to the official ChrF (word_order=0) used here.

The big jumps come from decoding, not just the model: ChrF-based Minimum-Bayes-Risk decoding (sample N candidates, pick the one maximizing mean pairwise ChrF over the deduplicated candidate pool) adds ~+2.4 ChrF with no extra training, and pooling candidates with a diverse NLLB-1.3B reaches 44.97.

Training

A two-stage 16-bit LoRA fine-tune of unsloth/Qwen3.5-9B (Unsloth; no 4-bit quantization; bf16; decoder-only, with minimal native Quechua pretraining), then merged to 16-bit weights. Both stages: LoRA r=256, α=512, dropout=0 over all 7 attention/MLP projections; adamw_8bit, weight-decay 0.01, warmup-ratio 0.05.

• Chanka corpus (1,929 pairs): 1,042 reviewed judicial-manual pairs + 503 manual-glossary entries + 349 Benito-2018 dictionary entries + 35 simple terms (deduped, leakage-filtered against the held-out eval split).
• Prompt: chat template, system "Eres un traductor profesional español-quechua", instruction "Traduce del español al quechua chanka…", reasoning disabled, loss on response only.
• Merge: Unsloth save_pretrained_merged (merged_16bit, no α-rescaling) of the Chanka adapter (ckpt-615) → this model.
• Not shipped here: a planned compact-mixed self-verification stage and a LoRA-α sweep were not part of this evaluated model (the SOTA number is the two-stage merge above).
• Zero leakage: the AmericasNLP 2021 test set was held out throughout (0/1003 overlap verified). Note the in-domain manual eval is inflated (glossary shares the manual the eval is drawn from); the AmericasNLP test (40.55) is the clean number.

Usage

from transformers import AutoTokenizer
from vllm import LLM, SamplingParams

tok = AutoTokenizer.from_pretrained("Thermostatic/rosettia-quy-v30b-9b-merged")
llm = LLM(model="Thermostatic/rosettia-quy-v30b-9b-merged", dtype="bfloat16")

SYSTEM = "You are a careful Spanish→Quechua (Chanka/Ayacuchana variety) translator."
INSTR  = "Traduce el siguiente texto del español al quechua chanka. Responde solo con la traducción."
src = "Mis abuelos eran una pareja muy cariñosa."
msgs = [{"role":"system","content":SYSTEM},
        {"role":"user","content":f"{INSTR}\n\nEspañol: {src}"}]
prompt = tok.apply_chat_template(msgs, tokenize=False, add_generation_prompt=True, enable_thinking=False)
out = llm.generate([prompt], SamplingParams(temperature=0.0, max_tokens=128,
                                            stop=["<|im_end|>","<|endoftext|>"]))
print(out[0].outputs[0].text.strip())

For best quality, use ChrF-MBR decoding (sample 32–64 candidates at T≈0.5, pick the consensus by mean pairwise sentence-ChrF over the deduplicated pool).

Intended use & limitations

Research-grade MT for a low-resource, agglutinative language. Outputs should be reviewed by speakers before any consequential use. Quality varies with domain and sentence length; named entities and Spanish loanwords are the most error-prone.

Authors & contributions

A two-person SomosNLP hackathon project:

• Estefanía Espinosa Fernández — data curation, and the initial Qwen3.5 LoRA experiments behind this model (comparing DoRA, rsLoRA and LoRA, and exploring data mixes).
• Irving Ernesto Quezada Ramírez (irvingernesto.com) — the subsequent modeling and the broader system this model is part of (synthetic distillation, the NLLB pipeline, GSPO RL, decoding/ensembling, evaluation, and release).

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