[FonaTech]

Summary

Adds models/chronos-enzh-50m-sft/, the packaged 50M Chronos v3 English → Chinese model from the 69-hour M5 Max run, and links it from the root READMEs.

• Contents: config.json, the model's own 16k tokenizer, the trained prompt (prompts.json), decoding settings (generation_config.json), the driver translate.py (CLI + Python API, pinned to Project_Chronos==0.3.2a0), the evaluation scripts, and aggregate results.
• Weights: not in git. They are attached to release enzh-50m-sft-v1. translate.py --download fetches them and checks SHA-256 57c59a14…fab4b7.
• Boundary, enforced by the driver:
  • English → Chinese only.
  • One sentence per call, ≤ 128 tokens.
  • zh→en is refused. en→zh-only SFT erased that direction: FLORES+ BLEU 0.00 after SFT, against 17.15 for the pretraining checkpoint.
  • Repetition loops are cut, with a warning.
• Measured (MLX bf16, greedy):
  • FLORES+ devtest: BLEU 33.48 / chrF 29.52
  • WMT22: 32.60 / 28.83
  • WMT24++ (sentence split): 26.89 / 24.43
• Terms: NOTICE limits the weights to research use. The training-corpus licences are unknown. The code is Apache-2.0.
• Packaging: MANIFEST.in prunes models/, so the PyPI sdist does not change.

Test plan

• ruff==0.17.4 check . and flake8 --select=E9,F63,F7,F82 are clean.
• The local CI simulation (no MLX, CPU torch 2.10.0, Python 3.11) passes: 262 passed / 49 skipped on main, 408 passed / 67 skipped on dev.
• tests/test_model_enzh_50m_sft.py (13 tests) checks:
  • refusal of zh, mixed-CJK, Japanese, Russian and empty input
  • that abbreviations and list numbers are not split
  • the loop guard
  • --direction zh_en exiting with code 2
  • the shipped settings
• With the exported weights, the driver reproduces the training run's FLORES+ BLEU 33.48.
• After the release: translate.py --download in an empty directory, plus one translation.

🤖 Generated with Claude Code