johnbean393/chiboard-1-t1

Chiboard-1 T1

Full fine-tune of LiquidAI/LFM2.5-1.2B-Base for Chinese pinyin-to-Hanzi conversion and revision. This is the 1.2B teacher initialization (T1).

Training exposure: all plain SFT rows once plus all revision SFT rows twice, concatenated and globally shuffled before one training pass.

Prompt format:

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

The tokenizer adds exactly one BOS. Loss is computed on the entire target plus <|im_end|> and nowhere in the prompt.

Exposure summary

{
  "eval": {
    "plain_eval_rows": 25000,
    "plain_split": "dev",
    "revision_eval_rows": 25000,
    "revision_split": "validation",
    "total_eval_rows": 50000
  },
  "revisions": {
    "model": "f6a5d174bc3e52bd0df245d69133f9930b4828d8",
    "plain": "54f9a912fce5f60df4918f83b3d9f0b5595f47b4",
    "revision": "62ffeead7a381d4d88d640a44531e3f3cbb9996e"
  },
  "train": {
    "effective_train_rows": 78233116,
    "logical_mixture_rows": 78233116,
    "max_train_samples": 0,
    "plain_row_exposure_share": 0.6970899382302502,
    "plain_source_rows": 54535518,
    "revision_repeats": 2,
    "revision_row_exposure_share": 0.30291006176974977,
    "revision_source_rows": 11848799,
    "shuffle_seed": 20260710
  }
}

Chiboard evaluation

• Harness: greedy decoding, population-weighted stratified sample (seed 20260714)
• Model revision: 4086865d8813a01909a420579ee7b15821bf80b0
• Evaluated: 2026-07-14
• Predictions: johnbean393/chiboard-1-eval/chiboard-1-t1-20260714

Split	Rows (population)	Exact match	CER	EOS rate	Empty rate
plain / dev	100,000 (1,139,609)	60.50%	0.1249	99.43%	0.62%
revision / dev	40,000 (60,305)	62.44%	0.0575	99.99%	0.00%

Aggregate metrics are population-weighted estimates from the deterministic base sample; exact match preserves whitespace and punctuation.