pantogloss-500-en-v7

Pantogloss 500-to-English successor v7

This repository contains the full-precision FP32 v7 Pantogloss model. V7 starts from the released v6 successor and applies a bounded conversational Spanish-to-English fine-tune. It retains the same architecture, tokenizers, many-to-English direction, and model-family language coverage. It improves an existing Spanish translation capability; it does not add a new language.

The weights are hosted separately from the Python wheel. This full-size model is an explicit opt-in model; the compact Pantogloss model remains the package default.

Usage

from pantogloss import Translator

translator = Translator.from_pretrained("pantogloss-500-en-v7", device="auto")
print(translator.translate("¿Cómo está usted?"))

Training

V7 uses 4,000 balanced Fisher/CALLHOME Spanish-to-English training examples prepared from the Joshua Decoder corpus project. Training-source utterances overlapping held-out Spanish sources were removed. The released candidate uses rank-8 LoRA on decoder cross-attention query/value projections only, followed by a 0.90 adapter merge into ordinary RTG weights and conversion back to the Pantogloss TensorFlow/Keras format.

Corpus text, evaluation text, and PyTorch optimizer state are not included in this repository. See V7-ATTRIBUTION.md for source attribution and preparation details.

Quality evidence

Against the released v6 parent under matched greedy decoding, the selected candidate produced these chrF improvements:

Candidate and parent produced zero runtime failures on every listed split. The two confirmation splits were opened once after checkpoint selection and the multilingual development gate were frozen.

On the pinned 8,250-example, 50-language FLORES+ development suite, v7 changed aggregate chrF by +0.217 and BLEU by +0.278 versus v6, with zero failures and no language below the frozen −0.5 chrF retention floor. Twelve languages had small negative chrF deltas. The largest were Indonesian −0.473, Slovenian −0.431, Danish −0.306, and Zulu −0.184. These observed declines are limitations even though the aggregate score improved.

“500-to-English” describes model-family training provenance, not uniform quality across 500 languages. The multilingual evaluation covers 50 languages.

Intended use and limitations

This model is intended for research and general-purpose text translation into English. It is not an automatic speech recognizer: Fisher/CALLHOME transcript text was used for fine-tuning, not audio. Conversational Spanish improvements may not transfer equally to every dialect, register, acoustic transcript, or general domain.

• Quality varies by language, domain, script, and sentence.
• The model does not detect the source language or provide calibrated confidence.
• Inputs can be mistranslated, shortened, hallucinated, or repeated.
• It has not been validated for safety-critical, medical, legal, or other high-stakes decisions. Users should review consequential translations.
• Automated metrics do not replace native-speaker review.

Provenance and license

The parent model descends from the RTG many-to-English model described by Gowda, Zhang, Mattmann, and May, ACL-IJCNLP 2021 System Demonstrations. The v7 model artifact is licensed under the Apache License, Version 2.0. Training sources retain their own terms and required attribution; see V7-ATTRIBUTION.md. Source creators, corpus contributors, LDC, and speakers do not endorse Pantogloss.