johnbean393/chiboard-1-t2-preview-0713

Chiboard-1 T2 Preview 0713

chiboard-1-t2-preview-0713 is a 1.2B-parameter Chinese pinyin input-method teacher derived from johnbean393/chiboard-1-t1. It is the first T2 experiment to beat the immutable T1 baseline on all six release metrics: exact match and character error rate (CER) for plain input, revision input, and hard-ambiguity rows.

This is a preview research checkpoint. The gains are positive but small; use the exact prompt format below and validate on your own input distribution.

Training

• Base: johnbean393/chiboard-1-t1 at commit 4086865d8813a01909a420579ee7b15821bf80b0
• Objective: DPO (beta=0.10) plus chosen-completion NLL anchoring (rpo_alpha=2.0)
• Update locality: final two transformer blocks and final norm
• Schedule: 130 optimizer steps, effective batch size 128, learning rate 2e-6, BF16
• Data: 50,000 deterministic train-only pairs: 19,887 clean-input, same-length, single-substitution preference pairs; 25,113 plain null preservation anchors; and 5,000 revision null preservation anchors
• Labeling: no LLM judge, no model judge, and no human review
• Diagnostic result: 56.66% active-pair reward accuracy, +0.01818 unscaled preference gain, chosen NLL improved by 0.000237, and 1.889% of broadly sampled trainable values changed

The active pairs were deterministically filtered from the audited johnbean393/chiboard-1-dpo-v2 lineage. All source datasets were pinned to immutable commits, and no development or test examples entered training.

Frozen evaluation versus T1

The release gate uses the same frozen, paired 2,500-row plain sample and 2,500-row revision sample for both models. Hard ambiguity is the combined 3,560-row subset containing ambiguity terms. Population weights are retained.

Cohort	T1 EM	T2 EM	EM delta (pp)	T1 CER	T2 CER	CER delta (pp)
Plain	0.612014	0.612477	+0.04628	0.121848	0.121841	-0.00071
Revision	0.638948	0.640010	+0.10621	0.056157	0.056135	-0.00221
Hard ambiguity	0.585340	0.585620	+0.02803	0.128080	0.128070	-0.00096

Lower CER is better. The repository includes the evaluation reports, manifests, and release acceptance record.

Prompt format

Construct the model input as:

{committed_context}<|reserved_6|>{raw_pinyin}<|reserved_7|>{provisional_display}<|reserved_8|>

Then generate greedily. The training/evaluation configuration uses a maximum input length of 512 tokens and maximum total sequence length of 768 tokens.

Example loading

from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "johnbean393/chiboard-1-t2-preview-0713"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    torch_dtype="bfloat16",
    device_map="auto",
)

Limitations

• This checkpoint is specialized for Chiboard's structured pinyin-to-text interface, not general chat or unrestricted text generation.
• The strict six-metric win is narrow, especially for plain and hard-ambiguity CER. Treat it as a preview and preserve the included evaluation protocol when comparing descendants.
• The evaluation is for Simplified Chinese distributions represented by the pinned Chiboard datasets; other domains, dialects, noisy keyboards, and Traditional Chinese may behave differently.
• Exact output depends on the reserved-token prompt contract. Ordinary chat prompts are out of distribution.

Lineage

Training seed: 20260711. Dataset-construction seed: 20260713. Tokenizer SHA-256: 516b5b72266074897f80b3bee95d1b8b74497dfc5e58d1be289db107c47b7c99.