johnbean393/chiboard-1-t2-dpo-exp051-step70
johnbean393/chiboard-1-t2-dpo-exp051-step70 is a 1.2 billion parameter language model, an experimental checkpoint from the Chiboard T2 series. This model was trained using local-DPO for 70 updates, focusing on improving specific metrics. It is primarily intended for reproducibility and error analysis, particularly for studying the trade-off between plain/hard-ambiguity gains and revision exact-match retention.
Loading preview...
Overview
This model, johnbean393/chiboard-1-t2-dpo-exp051-step70, is an experimental 1.2 billion parameter checkpoint from the Chiboard T2 series, derived from the johnbean393/chiboard-1-t2-preview-0713 model. It underwent 70 updates using a local-DPO training method with specific parameters, including four trainable final transformer blocks and FP32 next-token decision logits. The training incorporated 1,484 deterministic, train-only on-policy corrections without LLM judge or human review.
Key Characteristics & Performance
- Experimental Checkpoint: This is an unpublished, high-movement local-DPO checkpoint, made public for reproducibility and error analysis rather than as a promoted release.
- Targeted Training: Focused on improving specific metrics, particularly in 'Plain' and 'Hard ambiguity' slices, showing gains of +0.07133 EM and +0.02318 EM respectively compared to the accepted 0713 checkpoint.
- Known Regression: Despite improvements in some areas, the model experienced a significant regression in 'Revision' exact match by -0.49107 percentage points, leading to its non-promotion.
- Prompt Format: Utilizes a specific prompt format:
<|startoftext|>{committed_context}<|reserved_6|>{raw_pinyin}<|reserved_7|>{display}<|reserved_8|>, with generation stopping at<|im_end|>.
Intended Use Cases
- Research and Analysis: Ideal for researchers studying the trade-off between different performance metrics, specifically plain/hard-ambiguity gains versus revision exact-match retention in DPO training.
- Error Analysis: Useful for understanding the impact of specific DPO training configurations on model behavior and identifying areas of regression.
Limitations
- Not for Deployment: Due to the material regression in revision exact match, this checkpoint is not recommended for general deployment. Users seeking validated overall behavior should use the
johnbean393/chiboard-1-t2-preview-0713checkpoint instead.