pantogloss-500-en-v6-compact

Pantogloss 500-to-English v6 Compact

This release stores the released Pantogloss v6 fine-tune in a smaller FP16/INT5 format. Pantogloss 0.23.0 selects this model as its recommended default. The staged, checksummed runtime and legal artifacts total 564,137,246 bytes (538.0 MiB). Model weights are separate from the Python wheel.

The model uses the same architecture, source and target tokenizers, and many-to-English task as the full-precision v6 release. Eight existing source languages were targeted during v6 fine-tuning; this does not add new source languages. “500-to-English” describes model-family provenance, not uniformly validated quality across 500 languages. The evaluations below cover 50.

Usage

from pantogloss import Translator

translator = Translator.from_pretrained(
    "pantogloss-500-en-compact",
    device="gpu",
    packed_source_runtime="reconstruct",  # tested bulk CUDA path
)
print(translator.translate("Comment allez-vous ?"))

For Apple Metal, use packed_source_runtime="gather-sharded". The generic auto mode currently favors fast startup and uses gather-sharded on both backends. Explicit one-time reconstruction was faster for sustained CUDA inference but took longer to load; it is not the tested Metal path.

Conversion and provenance

The released v6 FP32 Keras weights were rounded to IEEE FP16. All non-source weights remain FP16. The source embedding uses symmetric five-bit storage per row with 2,800 exact FP16 overrides: 2,100 selected by INT5 reconstruction error and 700 by a language-balanced, source-only calibration rule. The calibration data are a pinned FLORES+ dev slice; quality was checked on disjoint devtest records. No translation references or test outputs were used to choose override rows. The manifest pins and checksums each conversion stage and every distributed artifact. No further gradient updates were made.

Evaluation

On a matched 8,250-example, 50-language FLORES+ confirmation corpus with greedy decoding and zero failures:

The direct CUDA comparison with the current Compact default found small chrF declines in seven of the 50 languages, led by Arabic (-0.189) and Lao (-0.161). The largest gains were Nepali (+4.107), Khmer (+2.278), and Hausa (+2.266). Aggregate gains do not mean every sentence or language improves.

Matched warmed throughput retained 96% and 87% of v6 FP16 on CUDA at batch 1 and 16 with explicit reconstruction. On M3 Metal, gather-sharded lookup retained 94% and 85% at those batch sizes. Peak process RSS was approximately 51% of matched FP16 on CUDA and 65%/61% on Metal. The full mixed-length Metal quality run at matched batch 128 retained 99% of FP16 throughput; these are different workloads and should not be conflated.

Limitations and redistribution

Quality varies by language, script, and domain. The model does not detect source language or provide calibrated confidence. It is not intended for unreviewed medical, legal, safety-critical, or other high-stakes use.

The model artifact is licensed under Apache-2.0. NOTICE credits the RTG parent model and original authors; V6-ATTRIBUTION.md lists the fine-tuning data sources. Corpus text and evaluation examples are not distributed with the model.