johnbean393/chiboard-1-eval

Chiboard 1 evaluation

This repository is the frozen, paired development evaluation for Chiboard 1 Chinese IME models. It contains greedy SGLang predictions for the T1 campaign baseline and the M0 and S1 reference checkpoints.

The current campaign baseline is johnbean393/chiboard-1-t1, not T2-preview-0713.

Campaign 20260714

• Seed: 20260714
• Frozen sample: 100,000 plain dev rows + 40,000 revision dev rows
• Plain population: 1,139,609 rows
• Revision population: 60,305 rows
• Sampling: deterministic, stratified, population-weighted; no targeted ambiguity extras
• Decoding: greedy, bfloat16, one BOS, Chiboard EOS, shuffled request order
• Engine: SGLang 0.5.9, native model implementation, default attention backend (resolved to FA3 on the campaign H100)
• Model context limit used by the harness: 33,024 tokens (MAX_INPUT_TOKENS=32768 plus generation allowance)

Performance

EM is higher-is-better; CER is lower-is-better. Values below are population-weighted estimates from the common frozen sample. Full register, length, noise, context, source, case, cadence, and ambiguity slices are available in each linked report.md.

Run	Model revision	Plain EM	Plain CER	Revision EM	Revision CER
T1	4086865d8813a01909a420579ee7b15821bf80b0	60.50%	0.1249	62.44%	0.0575
M0	0777e99abb4914f42cfebee33626f7c36638c671	50.39%	0.1447	36.14%	0.1337
S1	ac4e8c3885ef1a0ad5b3f90a1e4c5639ae669edd	54.84%	0.1341	54.68%	0.0774

Dataset Viewer

The Viewer exposes three configurations: t1 (default), m0, and s1. Each configuration contains a 100,000-row plain split and a 40,000-row revision split. The typed Parquet view under data/ mirrors the prediction JSONL artifacts; manifests and detailed reports remain in their campaign folders.

Repository layout

Each campaign folder contains:

• run_manifest.json: model and dataset commits, seed, limits, split mapping, sampling, and generation settings.
• predictions-plain-dev.jsonl: 100,000 prediction rows.
• predictions-revision-dev.jsonl: 40,000 prediction rows.
• report.json: machine-readable aggregate and slice metrics.
• report.md: rendered metrics, slices, and highest-CER failures.

Prediction rows include the stable example key, input, target, prediction, population sample weight, strata metadata, exact-match/CER fields, EOS status, and generation timing.

All three runs use identical (key, target, sample_weight) sequences. The validated frozen-sequence hashes are:

• Plain: 96dd4e7157a71c87c64ca14bad40416434d2e644884bee994ef3352c6cbe9ab7
• Revision: cd6383d99ed8b9bea47e99522a4682ca6414d9197198c3bd0ace655cb4ac2695

Dataset commits used for sampling:

• Plain, johnbean393/chiboard-1-sft: 54f9a912fce5f60df4918f83b3d9f0b5595f47b4
• Revision, johnbean393/chiboard-1-revision-sft: 62ffeead7a381d4d88d640a44531e3f3cbb9996e

Run the evaluation

Run from the Chiboard source repository, whose scripts/ directory contains the evaluation harness and launchers. Use a single H100-class GPU and keep caches on a volume with ample free space.

1. Install the campaign runtime

python3 -m venv /workspace/venv-sgl
source /workspace/venv-sgl/bin/activate
pip install \
  'sglang==0.5.9' \
  'transformers==4.57.1' \
  'datasets==5.0.0' \
  'huggingface_hub==0.36.2' \
  accelerate

export HF_HOME=/workspace/hf
export HF_TOKEN=YOUR_WRITE_TOKEN
mkdir -p /workspace/chiboard_eval

The token needs read access to the model and source datasets. It needs write access to johnbean393/chiboard-1-eval only when Hub synchronization is enabled.

2. Reproduce T1 and the references

cd /workspace/chiboard

PYTHON=/workspace/venv-sgl/bin/python \
MODEL=johnbean393/chiboard-1-t1 \
SEED=20260714 \
MAX_INPUT_TOKENS=32768 \
bash scripts/run_reference_evals_h100.sh

This performs three sequential 140,000-row passes:

1. T1, which defines /workspace/chiboard_eval/chiboard-1-t1-20260714.
2. M0 on the frozen T1 sample.
3. S1 on the same sample.

The launcher writes each engine chunk locally and uploads it to this dataset repository. Re-running the identical command restores uploaded manifests and prediction files, then resumes incomplete rows. Keep the seed, model IDs, dataset commits, limits, decode order, and engine settings unchanged when resuming.

Set PUSH_README=0 if the model cards should not be updated after reference runs. Set HUB_REPO=none for local-only output.

3. Evaluate a candidate on the frozen sample

First download the sample of record if it is not already local:

hf download johnbean393/chiboard-1-eval \
  --repo-type dataset \
  --include 'chiboard-1-t1-20260714/*' \
  --local-dir /workspace/chiboard_eval

Then decode the candidate:

PYTHON=/workspace/venv-sgl/bin/python \
MODEL=johnbean393/YOUR_CANDIDATE \
SEED=20260714 \
MAX_INPUT_TOKENS=32768 \
FROZEN_SAMPLE_DIR=/workspace/chiboard_eval/chiboard-1-t1-20260714 \
OUTPUT_DIR=/workspace/chiboard_eval/exp-YOUR_ID \
bash scripts/run_eval_sgl_h100.sh

The default DECODE_ORDER=shuffled is required for statistically valid partial looks.

4. Gate the candidate

During decoding, inspect a partial verdict:

python scripts/gate_eval_ci.py \
  --baseline /workspace/chiboard_eval/chiboard-1-t1-20260714 \
  --candidate /workspace/chiboard_eval/exp-YOUR_ID \
  --allow-partial

After 100% coverage, write the final verdict:

python scripts/gate_eval_ci.py \
  --baseline /workspace/chiboard_eval/chiboard-1-t1-20260714 \
  --candidate /workspace/chiboard_eval/exp-YOUR_ID

PASS requires the combined-utility 95% bootstrap CI lower bound to exceed zero and no non-inferiority guard violation. Do not change gate thresholds, margins, seeds, or cohorts after looking at candidate results.

5. Run harness tests

python -m pytest \
  tests/test_evaluate_chiboard_sgl.py \
  tests/test_gate_eval_ci.py \
  tests/test_check_eval_parity.py \
  tests/test_push_eval_to_readme.py \
  -q

Validation policy

For campaign 20260714, native SGLang output with its default attention kernel is the source of truth. Hugging Face/SGLang byte parity is not a release gate for these stored artifacts. Structural validation still requires complete row counts, pinned revisions, reports for both kinds, and identical frozen keys, targets, and sample weights across every model folder.

Intended use

Use this repository for paired Chiboard model comparison, CI gating, slice analysis, and reproducible candidate screening. It is an evaluation artifact, not a general-purpose training corpus. Preserve the campaign look budget and use a fresh seed for a new campaign.