This repository contains the model artifacts for the Pantogloss
TensorFlow/Keras many-to-English translation package. The model remains separate
from the Python distribution, so its approximately 2.18 GB of weights are never
included in the wheel.
Pantogloss 0.8.0 provides named decoding presets. fast is the default greedy
decoder; quality selects beam size 4 and length penalty 0.6; custom supports
explicit beam_size and length_penalty values. Greedy remains the default
because beam-4 was approximately 16 times slower in the full evaluation below.
Install pantogloss[cuda] for NVIDIA acceleration on Linux or
pantogloss[metal] for Apple Silicon. Pantogloss 0.3.0 and later resolve this
public repository anonymously. Cached and explicit tokens remain supported for
private or gated repositories.
Language quality metadata
Pantogloss 0.10.0 can attach versioned benchmark evidence when the caller
supplies an ISO 639-1 or BCP-47 source-language hint:
python
1result = translator.translate_detailed(2"ສະບາຍດີ",3 source_language="lo",4 include_warnings=True,5)6print(result.language_quality.to_dict())7print([warning.to_dict()for warning in result.warnings])
The same metadata is available as JSON Lines from the CLI:
pantogloss translate --json --source-language lo --include-warnings "ສະບາຍດີ"
Pantogloss does not detect the language. It normalizes a caller-supplied BCP-47
hint to its primary code, such as pt-BR to pt. The current
flores-plus-v4.6-greedy-v1 catalog contains aggregate greedy chrF measurements
for 50 languages over an immutable 8,250-example FLORES+ selection:
measured-higher: chrF 60 or above;
measured-mixed: chrF from 45 through 59.99;
measured-limited: chrF below 45; and
unmeasured: no row in this catalog, not a claim of unsupported input.
These tiers describe benchmark evidence for the pinned model and corpus. They
are not sentence-level confidence, calibrated uncertainty, a quality guarantee,
or a language-support declaration. Opt-in warnings are conservative review
signals for limited or unmeasured evidence and observable output risks such as
source copying, retained source script, repetition, empty output, and extreme
length ratios.
The complete measured-language table, provenance, and interpretation guidance
are maintained in the
Language Quality Catalog.
Validation
All 308 learned PyTorch tensors map to 307 Keras variables because the target
embedding and output projection are tied. CPU greedy token IDs match the archived
RTG implementation across a ten-language, six-script batch. The maximum absolute
difference in inference-critical tensors is 1.24e-5. With the original beam size
4 and length penalty 0.6, the complete decoded four-best candidate sets match on
all ten examples; one near-tied example changes top-candidate rank because of
framework floating-point ordering.
CUDA placement and inference have been validated on an NVIDIA GeForce RTX 3080 Ti
Laptop GPU. Metal placement and inference have been validated on an Apple M3 Max
with 128 GB unified memory, macOS 26.5.2, Python 3.12.9, and TensorFlow 2.18.1.
The first model variable was confirmed on /GPU:0, and translation matched CPU.
Pantogloss 0.2.0 uses graph-compiled greedy decoding and decoder attention
caches. Warm throughput for batches 1, 8, 16, and 32 was 11.6, 62.1, 96.8, and
143.0 sentences/second on Linux CPU; 14.2, 94.9, 160.6, and 330.0 on CUDA; and
3.83, 29.15, 60.20, and 115.28 on Metal. These figures use one repeated short
input and characterize runtime behavior, not translation quality.
Full multilingual evaluation
Pantogloss 0.8.0 was evaluated on a deterministic 8,250-example, 50-language
selection from FLORES+ 4.6. Greedy decoding scored BLEU 31.48 (95% paired-
bootstrap interval 31.07–31.96), chrF 57.33 (56.98–57.69), and COMET 0.82975.
Beam-4 with length penalty 0.6 scored BLEU 32.42 (32.00–32.86), chrF 58.03
(57.70–58.38), and COMET 0.83528. Both modes had zero failures, empty outputs,
or literal unknown-token outputs.
Beam improved BLEU by 0.94, chrF by 0.70, and COMET by 0.00553, but sustained
only 1.22 sentences/second versus greedy's 19.89 on the RTX 3080 Ti Laptop GPU.
The intervals use 1,000 deterministic paired-bootstrap samples and seed 1729.
COMET used Unbabel/wmt22-comet-da at immutable revision
2760a223ac957f30acfb18c8aa649b01cf1d75f2, unbabel-comet 2.2.7, and
PyTorch 2.7.1 with CUDA 12.8.
COMET's declared coverage does not include Igbo, Yoruba, or Zulu, representing
300 examples; their COMET scores are unreliable under the metric model's own
coverage warning. FLORES+ is Wikipedia-derived and does not establish equal
quality across domains. The public aggregate report and exact reproduction
metadata are maintained in the Pantogloss code repository; gated source,
reference, and translation text are not redistributed.
Provenance
The weights were converted from the massively multilingual model described by
Thamme Gowda, Zhao Zhang, Chris A. Mattmann, and Jonathan May in Many-to-English
Machine Translation Tools, Data, and Pretrained Models, ACL-IJCNLP 2021 System
Demonstrations, DOI 10.18653/v1/2021.acl-demo.37.
Coverage and translation quality vary by language and domain. “500-to-English”
describes training provenance, not equal quality across every language.
Intended use and limitations
This model is intended for research and general-purpose translation of text into
English. It has not been evaluated for safety-critical, legal, medical, or
high-stakes decisions. Users should review translations before relying on them.
Training coverage does not imply reliable language identification or equal
quality for every language, script, dialect, or domain.
Inputs outside the training distribution can be mistranslated, shortened,
hallucinated, or returned as low-quality English.
The current model does not return confidence, detected language, citations,
or calibrated uncertainty. Pantogloss quality tiers are aggregate benchmark
metadata and must not be treated as confidence.
The parity corpus validates checkpoint conversion, not broad translation
quality, fairness, or safety.
Source data may contain biases that can be reproduced in generated text.
Artifact integrity
Pantogloss pins model version 0.1.0 to immutable Hugging Face revision
250fc3b4122d79ac0734b28b368d2c1d68f72f7e. manifest.json records byte sizes
and SHA-256 digests for every runtime artifact. The two weight shards are:
