nllb-sango-finetuned-600m-v2-3 — superseded experiment, published for the record

Use MEYNG/nllb-sango-finetuned-600m-v2-2 instead. This checkpoint is the v2.3 run of MEYNG's five-experiment Sango↔French program. It is beaten by v2-2 in both directions and is kept public only so the negative result stays checkable.

This model was uploaded on 2026-07-17 with an auto-generated placeholder card and no licence declaration. This card replaces it. Nothing about the weights has changed.

What this run tested

v2.3 changed exactly one variable from the v2.2 recipe: it added native-verified vocabulary, oversampled ×5, to the cleaned training corpus. The hypothesis was that a small amount of high-quality, human-checked lexical data would improve out-of-distribution generation.

It did not. From the five-experiment log published in the v2-2 model card:

Run	One variable changed	FR→SAG Δ vs baseline	SAG→FR Δ vs baseline
v2.2 (recommended)	lr 2e-4→1e-4, OOD-selected checkpoint	−1.71	+1.25
v2.3 (this model)	+ native-verified vocabulary ×5	−2.22	+1.11

v2.3 is worse than v2.2 on both directions. The conclusion MEYNG drew from it — that small high-quality vocabulary injection does not improve OOD generation — is the reason this run exists and the reason it did not become the recommended model.

Those deltas are chrF against the untrained facebook/nllb-200-distilled-600M baseline, as recorded in that log. They correspond to the SIB-200 column: the v2.2 row in the same table (−1.71 / +1.25) matches v2-2's published SIB-200 deltas exactly. Per-benchmark FLORES+ figures for this run have not been published, and no evaluation artifact for v2.3 exists in MEYNG's results directory. Do not quote a FLORES number for this model.

Direction

As with every model in this series, the honest summary is direction-dependent: SAG→FR improves on the baseline, FR→SAG does not. For French→Sango, use facebook/nllb-200-distilled-600M directly. For Sango→French, use v2-2, not this model.

Model details

• Base model: facebook/nllb-200-distilled-600M (Meta AI, NLLB-200). Architecture M2M100ForConditionalGeneration, d_model 1024, 12 encoder + 12 decoder layers, vocab 256,206.
• Method: LoRA fine-tune, merged to standalone weights (no peft needed at inference).
• Requires transformers>=4.40,<5.0 — v5 breaks the NLLB tokenizer prefix.
• Language tokens: NLLB conditioning via forced_bos_token_id; sag_Latn / fra_Latn.

Usage

from transformers import AutoModelForSeq2SeqLM, AutoTokenizer  # transformers>=4.40,<5.0

model_id = "MEYNG/nllb-sango-finetuned-600m-v2-3"  # prefer ...-v2-2
tok = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForSeq2SeqLM.from_pretrained(model_id)

def translate(text, src="sag_Latn", tgt="fra_Latn"):
    tok.src_lang = src
    enc = tok(text, return_tensors="pt")
    out = model.generate(**enc,
                         forced_bos_token_id=tok.convert_tokens_to_ids(tgt),
                         max_length=160, num_beams=4)
    return tok.batch_decode(out, skip_special_tokens=True)[0]

Limitations

• Superseded. v2-2 is better in both directions. There is no use case where this model is the right choice over v2-2.
• FR→SAG trails the baseline out-of-distribution (−2.22 chrF): use the baseline for that direction.
• Domain skew: training data is predominantly religious/NGO-register web text; conversational and technical Sango are untested.
• "Verified" in MEYNG's vocabulary dataset means pipeline-verified, not native-speaker sign-off.
• No FLORES+ figures published for this run (see above).

Licence

CC BY-NC 4.0 — non-commercial use only.

This licence is inherited, not chosen. facebook/nllb-200-distilled-600M is released by Meta under CC BY-NC 4.0. These weights are a derivative work of it, so the NonCommercial term carries through.

Permitted: research, academic work, teaching, humanitarian and NGO deployment, public-sector and personal use. Not permitted: paid APIs or subscriptions, licensing to commercial entities, bundling into paid products, advertising-supported services.

Under Creative Commons' guidance the NonCommercial term restricts the use, not the user: a company may use these weights for genuinely non-commercial purposes, and a non-profit may breach the term by using them commercially.

Commercial use requires separate permission from Meta (for the base model) and from MEYNG (for the fine-tune). MEYNG cannot grant the first.

Attribution

This model is a derivative of NLLB-200 by Meta AI (facebook/nllb-200-distilled-600M, CC BY-NC 4.0). Redistribution or adaptation of these weights must credit both the base model and this fine-tune.

NLLB Team et al., No Language Left Behind: Scaling Human-Centered Machine Translation, 2022. arXiv:2207.04672

About MEYNG

MEYNG builds language AI infrastructure for under-resourced African languages, starting with Sango (Central African Republic, 5M+ speakers). Out-of-distribution gates before every release, failures documented and left public — this card is one of them. https://huggingface.co/MEYNG · https://meyng.com