Skip to content

chronos-enzh-50m-sft v1 (en→zh, research only)

Choose a tag to compare

View all tags
@FonaTech FonaTech released this 08 Nov 18:24
enzh-50m-sft-v1
e7400b5

Model weights only. This is not a Project_Chronos package release. The PyPI kernel stays at 0.3.0 (stable) / 0.3.2a0 (pre-release).

chronos-enzh-50m-sft is a 49.7M-parameter Chronos v3 recurrent-depth MoE (12.1M active per token). It was trained for 69 hours on one Apple M5 Max with MLX: pretraining on bilingual data, then English → Chinese SFT. The code, config, tokenizer, prompt and evaluation are in models/chronos-enzh-50m-sft/.

Boundary

  • English → Chinese only. Chinese → English is disabled. English-only fine-tuning erased that direction: on 200 FLORES+ devtest sentences, BLEU is 0.00 after SFT, against 17.15 for the pretraining checkpoint.
  • One sentence per call, at most 128 tokens. Joining sentences drops content. WMT24++ items over 128 tokens score BLEU 6.91 as one input and 28.62 when split into sentences.
  • Translation, not chat. Instructions inside the source are translated, not executed.
  • 50M parameters. Expect wrong quantities, literal idioms, omissions, destroyed URLs and code, and half-width punctuation.

translate.py enforces the first two rules and cuts repetition loops (3 of 1012 FLORES+ outputs).

Measured

Setup: MLX bf16, greedy decoding, Project_Chronos 0.3.2a0, sacrebleu 2.6.0, BLEU tokenize="zh".

Test set (en→zh) Scored / total BLEU chrF
FLORES+ devtest 1012 / 1012 33.48 29.52
WMT22 general (split into sentences) 2036 / 2037 32.60 28.83
WMT24++ en-zh_CN (split into sentences) 959 / 960 26.89 24.43

All evaluation sets were removed from the training data on both sides.

Use

cd models/chronos-enzh-50m-sft
pip install --pre "Project_Chronos[mlx]==0.3.2a0"   # or "Project_Chronos==0.3.2a0" without MLX
python translate.py --download       # fetches this asset, checks SHA-256
python translate.py "The meeting was moved to Friday."

Assets

File Size SHA-256
chronos-enzh-50m-sft.safetensors (bfloat16, tensors only) 99,470,698 bytes 57c59a141f3f75898268bbe2cda6d3218d514b229ad7c8e61fc02cc270fab4b7

Terms

Research use only. No commercial use. The model was trained on WMT19 zh-en, OPUS-100 en-zh and a WMT CSV export. Their licences are listed as unknown and have not been verified. You are responsible for checking that your use complies with the terms of the original data. The weights come with no warranty. Redistribution must keep the NOTICE and the SHA-256 above.

Assets 4
Loading