---
license: cc-by-sa-4.0
language:
- zh
tags:
- chiboard
- pinyin
- ime
- sft
- trl
base_model: LiquidAI/LFM2.5-350M-Base
datasets:
- johnbean393/chiboard-1-sft
---

# chiboard-1-m0

Bootstrap 350M SFT model for Chiboard mistake mining.

Prompt format:

```text
<|startoftext|>{committed_context}<|reserved_6|>{raw_pinyin}<|reserved_7|>{display}<|reserved_8|>{target}<|im_end|>
```

The model was trained with completion-only loss on `{target}<|im_end|>`.

Serve with exactly one `<|startoftext|>` token. Most runtimes add it automatically, so do not also embed it in the prompt string.

This is M0, a data-generation tool for typing-prefix replay; it is not the shipped final model.

## Training

- Base model: `LiquidAI/LFM2.5-350M-Base`
- Dataset: `johnbean393/chiboard-1-sft`
- Training layout: `mixed_packed`
- Max packed length: `4096`
- Effective batch: `10 x 6 = 60` packed rows per optimizer step
- Steps: `14746` / `14746`
- Final eval loss: `0.17373667657375336`
- Final eval mean token accuracy: `0.9519697650369391`
- Train runtime seconds: `3.506e+04`
- Packed LFM2 short-conv boundary isolation: `seq_idx` collator enabled

<!-- chiboard-eval:start -->
## Chiboard evaluation

- Harness: greedy decoding, population-weighted stratified sample (seed `20260714`)
- Model revision: `0777e99abb4914f42cfebee33626f7c36638c671`
- Evaluated: 2026-07-14
- Predictions: [`johnbean393/chiboard-1-eval/chiboard-1-m0-20260714`](https://huggingface.co/datasets/johnbean393/chiboard-1-eval/tree/main/chiboard-1-m0-20260714)

| Split | Rows (population) | Exact match | CER | EOS rate | Empty rate |
|---|---:|---:|---:|---:|---:|
| plain / dev | 100,000 (1,139,609) | 50.39% | 0.1447 | 99.44% | 0.19% |
| revision / dev | 40,000 (60,305) | 36.14% | 0.1337 | 99.94% | 0.00% |

Aggregate metrics are population-weighted estimates from the deterministic
base sample; exact match preserves whitespace and punctuation.
<!-- chiboard-eval:end -->