johnbean393/chiboard-1-m0
johnbean393/chiboard-1-m0 is a 350 million parameter instruction-tuned language model, bootstrapped from LiquidAI/LFM2.5-350M-Base, specifically designed for Chiboard mistake mining. This model is optimized for data generation in typing-prefix replay scenarios, focusing on predicting target text based on committed context, raw Pinyin, and display information. It achieves a final evaluation mean token accuracy of 95.19% and is intended as a data-generation tool rather than a final shipped model.
Loading preview...
Model Overview
johnbean393/chiboard-1-m0 is a 350 million parameter supervised fine-tuned (SFT) model, built upon the LiquidAI/LFM2.5-350M-Base architecture. Its primary purpose is to serve as a data-generation tool for Chiboard's mistake mining, specifically for typing-prefix replay. This model is not intended as a final user-facing product but rather as an intermediate step in data creation.
Key Capabilities
- Contextual Prediction: The model processes a unique prompt format including committed context, raw Pinyin, and display information to predict a target string.
- Completion-Only Loss: Training focused on optimizing for the target completion, ensuring high accuracy in generating the desired output.
- High Token Accuracy: Achieved a final evaluation mean token accuracy of 95.19% during training.
Training Details
The model was trained on the johnbean393/chiboard-1-sft dataset using a mixed_packed layout with a maximum packed length of 4096 tokens. It underwent 14,746 steps with an effective batch size of 60 packed rows per optimizer step. The training incorporated packed LFM2 short-conv boundary isolation with seq_idx collator enabled.
Evaluation Metrics
Evaluated using greedy decoding on a population-weighted stratified sample, the model demonstrated:
- Plain / Dev Split: 50.39% exact match and 0.1447 Character Error Rate (CER) on 100,000 rows.
- Revision / Dev Split: 36.14% exact match and 0.1337 CER on 40,000 rows.
These metrics highlight its performance in generating accurate text completions for its specialized data-generation task.