ajaxdavis/mobtranslate-migmaq-listuguj-v3-3

MobTranslate Mi'kmaq Listuguj v3.3

This is a noncommercial research LoRA for English-to-Mi'kmaq sentence translation in the Listuguj orthography. It is the best candidate from the MobTranslate v3.3 dialog-only experiment, selected prospectively across seeds 17, 42, and 73 and tested once on a sealed 133-sentence lesson-source split.

It is not a reliable general translator. It has not received fluent-speaker review, has not been tested across independent speakers, and makes material lexical, kinship, person/number, and inflection errors. Do not use its output as authoritative language, or for health, legal, educational-assessment, ceremonial, or other consequential communication.

Intended use

• Reproducible low-resource machine-translation research.
• Comparing decoder, adaptation, retrieval, and human-review methods.
• Producing explicitly labelled drafts for expert review.

Known dictionary queries should use deterministic dictionary lookup. This model reconstructed only 1 of 14,438 eligible dictionary records exactly in its full development census; neural lexical recall is not a dictionary service.

Usage

python -m pip install -r requirements.txt
python inference.py "Who is this?"

from peft import PeftModel
import torch
from transformers import AutoModelForSeq2SeqLM, AutoTokenizer

base_id = "ajaxdavis/mobtranslate-migmaq-listuguj-v1-base"
model_id = "ajaxdavis/mobtranslate-migmaq-listuguj-v3-3"
base_revision = "v1.0.0-rc1"
model_revision = "v3.3.0"
tokenizer = AutoTokenizer.from_pretrained(
    model_id,
    revision=model_revision,
    src_lang="eng_Latn",
    tgt_lang="mic_Latn",
)
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
base = AutoModelForSeq2SeqLM.from_pretrained(
    base_id,
    revision=base_revision,
    torch_dtype=torch.bfloat16,
)
base.resize_token_embeddings(len(tokenizer))
model = PeftModel.from_pretrained(
    base, model_id, revision=model_revision
).to(device).eval()
inputs = tokenizer("Who is this?", return_tensors="pt").to(device)
output = model.generate(
    **inputs,
    forced_bos_token_id=256204,
    max_new_tokens=192,
    num_beams=4,
    do_sample=False,
    no_repeat_ngram_size=3,
    repetition_penalty=1.1,
    length_penalty=1.0,
)
print(tokenizer.decode(output[0], skip_special_tokens=True))

The custom target token is mic_Latn at immutable tokenizer ID 256204. Inputs are unprefixed English sentences. The published generation contract is beam search with 4 beams, no sampling, a 3-token no-repeat n-gram guard, repetition penalty 1.1, and length penalty 1.0.

The evaluated runtime used BF16. Loading in float32 or changing hardware, library versions, batching, or decoder settings can change beam-search outputs.

Training

The selected rank-32 LoRA was trained for 600 optimizer updates at effective batch size 32 from the exact MobTranslate Mi'kmaq v1 checkpoint. Its 19,200 presentations comprised 11,520 source-attested Mi'gmaq Online dictionary-example sentences and 7,680 source-attested Listuguj lesson-dialog sentences. The schedule contains 6,247 unique source rows, 261,363 source tokens, 340,491 target tokens, and no truncated presentations. AdamW used learning rate 5e-5, 48 warmup updates, and seed 17. LoRA targeted q_proj, k_proj, v_proj, out_proj, fc1, and fc2 with rank 32, alpha 64, and dropout 0.05.

Evaluation

The fixed decoder was used for all comparisons.

Endpoint	Base	Step-matched retention	v3.3 candidate
Sealed chrF++ (133 rows)	22.3450	23.3204	28.1487
Sealed SacreBLEU	6.3032	5.6464	7.8084

Candidate minus retention chrF++ was +4.8284; the paired-row bootstrap 90% lower bound was +3.2820. A component bootstrap over 13 connected lesson components gave a 95% interval of [+2.7147, +8.0593]; 11 of 13 components were positive. Across the paired seed-17/42/73 development runs, mean dialog gain was +4.2962 chrF++ with a hierarchical 90% interval of [+2.8998, +5.7880].

The sealed gain is concentrated in dialogue: +6.1099 chrF++ over retention on 101 dialogue rows, versus +0.6003 on 32 vocabulary-container phrases. Of 133 sealed rows, 80 improved, 14 tied, and 39 worsened by sentence-level chrF++. Only 3 outputs exactly matched the single reference. These metrics come from one pedagogical repository and are not estimates of general population reliability.

Full machine-readable comparisons, prediction diagnostics, bootstrap settings, metric signatures, contracts, and hashes are under evaluation/ and provenance/. The adapter SHA-256 is 82d4c0d18ad14b2ae7c8a3684053815ebf54a402a953dd4c7d01e9e69307b687.

The adapter release tag is v3.3.0; it must be loaded over ajaxdavis/mobtranslate-migmaq-listuguj-v1-base at tag v1.0.0-rc1. A BF16 merged checkpoint was explicitly rejected because its beam-search generations differed from direct adapter execution on 3 of 6 fixed probes. The adapter is the artifact that was actually evaluated.

The staged base-plus-adapter package passed a separate six-input CPU-BF16 load smoke: all artifact and tokenizer hashes matched, all outputs were nonblank, target-prefixed, non-copies, and EOS-terminated. Exact cross-runtime text is not a release gate because beam outputs can change with batching, hardware, and numeric kernels; the sealed metrics above come only from the frozen GPU evaluation files.

Sources and license

• Base architecture: facebook/nllb-200-distilled-600M, CC BY-NC 4.0.
• Mi'gmaq Online material: CC BY-NC 4.0.
• Listuguj lesson material: CC BY 4.0, source commit c424e98c3d87c3890618fd63cdf5af7ad22b3009.

The combined checkpoint is released under CC BY-NC 4.0. Commercial use is not permitted under the inherited terms. Application-code licenses do not override the model or dataset licenses.

Release decision

Research publication preserves the exact evaluated base-plus-adapter execution. Homepage routing and production API use remain not authorized by this release. The deterministic dictionary route remains separate.