johnbean393/chiboard-1-t2-dpo-exp048-step148

TEXT GENERATIONConcurrent Unit Cost:1Model Size:1.2BQuant:BF16Context Size:32kPublished:Jul 13, 2026Architecture:Transformer Featherless Exclusive Cold

The johnbean393/chiboard-1-t2-dpo-exp048-step148 is a 1.2 billion parameter language model, derived from Chiboard T1, specifically trained for Chinese pinyin-to-Hanzi IME research. This model utilizes FP32 DPO sequence-score accumulation with a beta of 0.1 and a chosen-answer NLL/RPO anchor of 2.0, featuring a 32768 token context length. It serves as an experimental checkpoint for reproducibility and ablation studies in IME development, demonstrating improved EM and CER metrics over its T1 predecessor. Its primary application is in advancing research for Chinese input method editors.

Loading preview...

Chiboard-1 T2 DPO experiment 048 — step 148

This model, developed by johnbean393, is a 1.2 billion parameter experimental checkpoint from the Chiboard T2 series, designed for Chinese pinyin-to-Hanzi IME research. It was trained for one epoch using FP32 DPO sequence-score accumulation, a beta of 0.1, and a chosen-answer NLL/RPO anchor of 2.0, building upon the Chiboard T1 base.

Key Characteristics & Performance

  • Architecture: 1.2B parameter model, based on Chiboard T1.
  • Training Method: Utilizes conventional DPO (Direct Preference Optimization) with specific parameters for sequence-score accumulation.
  • Prompt Format: Employs the immutable Chiboard v2 prompt format, designed for pinyin-to-Hanzi conversion.
  • Evaluation: Demonstrated improvements over Chiboard T1 across all six point metrics (EM and CER) in frozen evaluations, including slices for Plain, Revision, and Hard ambiguity. It showed a combined utility of +0.04233 pp against the accepted johnbean393/chiboard-1-t2-preview-0713 checkpoint.

Intended Use Cases

  • IME Research: Primarily intended for research and ablation comparisons in Chinese pinyin-to-Hanzi Input Method Editors.
  • Reproducibility: Publicly available to ensure reproducibility of experimental results.
  • Ablation Studies: Useful for comparing different DPO configurations and their impact on IME performance.

Limitations

This model is an experimental checkpoint and not the recommended production release. Small metric differences may not be reproducible across different samples, decoding policies, prompt serializations, or tokenizers. For overall release quality, the accepted johnbean393/chiboard-1-t2-preview-0713 model is recommended.