pantogloss-500-en-v7-compact

Pantogloss 500-to-English v7 Compact

This is the compact TensorFlow/Keras release of Pantogloss v7 and the recommended Pantogloss default. It preserves the v7 conversational Spanish-to-English improvements in a 564,891,973-byte model-artifact payload (approximately 538.7 MiB). Model weights are distributed separately from the Python wheel.

The model uses the same architecture and tokenizers as the full v7 release. It translates into English and does not detect source language. “500-to-English” describes model-family provenance, not uniformly validated quality across 500 languages; the multilingual confirmation below covers 50 languages.

Usage

from pantogloss import Translator

translator = Translator.from_pretrained(device="auto")
print(translator.translate("Comment allez-vous ?"))

The model is also selectable explicitly as pantogloss-500-en-compact.

Compact representation

All non-source weights use IEEE FP16. The source embedding uses symmetric five-bit storage per row with 3,300 exact FP16 overrides: 2,100 selected by INT5 reconstruction error, 150 by Xhosa calibration-weighted reconstruction impact, and 1,050 by language-balanced calibration impact. Selection used only source text from a pinned FLORES+ dev calibration split. Confirmation used disjoint devtest records, and no confirmation references or outputs were used to rank override rows. No gradient updates were made during compaction.

Evaluation

On the 8,250-example, 50-language confirmation corpus, greedy decoding produced zero failures. Relative to full FP32 v7, compact v7 changed aggregate BLEU by -0.0535 and chrF by -0.0405. The largest per-language chrF decline was Xhosa at -0.7065. Relative to the previous compact default, aggregate chrF improved by 0.0168.

The compact model retained the full v7 conversational Spanish gains. It improved Fisher test chrF by 2.7454 and CALLHOME evltest chrF by 2.0286 over full v6. Across five conversational splits, its chrF loss relative to full v7 ranged from -0.0075 to -0.1282, with zero translation failures.

Validated runtime paths include an NVIDIA RTX 3080 Ti Laptop GPU with CUDA and an Apple M3 Max with Metal. Exact results and frozen gates are recorded in the Pantogloss repository under training/experiments/successor-v7-compact.

Limitations and redistribution

Quality varies by language, script, and domain. This model provides no source language detection or calibrated confidence and is not intended for unreviewed medical, legal, or other safety-critical use.

The model artifact is licensed under Apache-2.0. NOTICE credits the RTG parent model and original authors; V7-ATTRIBUTION.md records fine-tuning data sources. Training corpora and evaluation examples are not distributed here.