pantogloss-500-en-fp16

Pantogloss 500-to-English FP16

This is the released FP16 variant of Panto-500, a many-to-English translation model for Pantogloss. It is a deterministic IEEE-754 binary16 conversion of released float32 model version 0.1.0. The original chrismattmann/pantogloss-500-en repository remains the immutable numerical reference and Pantogloss default.

The FP16 model is intended for memory-constrained CUDA and Apple Metal use. It uses FP16 variables, activations, and attention caches. Beam log probabilities and accumulated search scores remain float32 to protect ranking stability.

Use

Pantogloss 0.11 or later includes the FP16 model in its catalog:

from pantogloss import Translator

translator = Translator.from_pretrained("pantogloss-500-en-fp16", device="gpu")
print(translator.translate("Comment allez-vous ?"))

The model downloads from Hugging Face on first use and is not included in the Pantogloss wheel. The repository is public and does not require authentication.

Validation

The complete greedy quality gate used the pinned FLORES+ 4.6 corpus: 8,250 examples across 50 languages. FP16 completed with zero failures, empty outputs, or literal unknown-token outputs.

FP16 changed 173 of 8,250 translations: sentence chrF classified 95 as improved, 78 as regressed, and 8,077 as unchanged. The largest per-language chrF decrease was 0.278 for Norwegian Bokmal. These scores characterize the pinned evaluation corpus, not every language, domain, or document type.

Three fresh-process CUDA repetitions measured approximately 49% lower TensorFlow allocator peak and 38.5% lower process RSS. On an Apple M3 Max, three paired fresh-process repetitions measured approximately 42.1% lower peak process RSS. M3 median throughput was +1.9% at batch 1 and -4.1% at batch 16. FP16 therefore offers a reliable storage and memory benefit; speed varies by backend and workload.

Greedy and beam-4 multilingual smoke tests, targeted weak-language tests, CUDA tests, and Apple Metal tests completed without new failures. Apple Metal greedy inference uses /GPU:0; beam decoding uses Pantogloss's validated CPU fallback.

See docs/fp16-experiment.md and the public-safe aggregate report in evaluation/reports/pantogloss-fp16-candidate-summary.json for detailed methodology, checksums, limitations, and reproducibility evidence. The gated FLORES+ source, references, and translations are not published.

Integrity and provenance

manifest.json declares precision: float16, released model version 0.1.0, the immutable source model version and manifest digest, every runtime artifact size and SHA-256, and the maximum observed weight-rounding error. All 307 Keras variables were assigned by stable path. The weight shard is about 1.08 GB, 49.98% smaller than the float32 weights.

License

The Pantogloss implementation and converted model are licensed under the Apache License, Version 2.0. See LICENSE and NOTICE.