State Hints V3 EN - teamswyg/laya-tools GitHub Wiki

State hints V3: actual warm-start training and research publication

한국어 · Feature guide · Earlier 100-case observations · Home

We actually continued training a small Go model that reads short work messages to propose displays for progress, completion reports and questions. Model publication and complete readback are finished. However, completion mistakes increased, so this remains a research candidate without automatic default activation or policy replacement.

This page records V3 training and publication, separately from work on an actual Riido reader. The results below come from a synthetic fixture reader. They do not verify live application reads, work mutations or completion of real tasks.

What was trained?

The model is a FP32 linear softmax classifier with 1,024 features and 8 intents. Its Go RSH artifact is 32,960 bytes. It does not generate sentences, and no Laya-weight, LoRA, distillation or 1.58-bit-model claim is made.

The 8 intents are question, blocker, reference, progress, completion_report, cancel_request, planned and unclear. The display experiment targets progress, completion reports and questions. A completion report is the author's assertion that work finished, not independently verified completion.

The parent is Go v0.2, pinned to SHA-256 ae6761dc81501b39ae17f29b48f0f0a8b35305fe11e297b9a787af0f9636df05. CE+AdamW actually ran on 400 new original synthetic messages starting from those weights. Optimizer state restarted.

Selected model condition Value
Selected arm warm_lr005
Epochs / batch / seed 40 / 32 / 1729
Learning rate / decay 0.005 / 0.001
New optimizer updates 520
Cumulative training steps 1,920 → 2,440
Calibration temperature 1.0
Confidence / margin 0.9 / 0.05 unchanged

Selection among 4 arms used validation NLL on 80 messages for the three display intents plus none. None is the probability sum of the other 5 intents. A separate 80-message calibration set selected the temperature. The model and conditions were saved before the 200-message test was first parsed. Selection or display criteria were not changed to improve the final test result.

What improved, and what regressed?

The parent and selected candidate were compared on the same new synthetic test of 200 messages.

Agreement with author annotations Parent v0.2 Research v0.3
Correct across all 8 intents 152/200 158/200
Progress correct 45/50 44/50
Completion-report correct 40/50 46/50
Question correct 24/50 36/50
Correct proposals / all proposals 57/57 90/94
Incorrect completion-report proposals 0 4

Overall agreement, question accuracy and the number of correct proposals increased. Progress accuracy fell by one, and 4 incorrect completion-report proposals appeared. Proposing more and making fewer wrong displays are different objectives. This trade-off remains visible, and default activation is withheld.

Full 8-intent confusion, probabilities, abstentions, failures and diagnostics remain in the exact first results. Live reads, mutations and state changes were zero. Synthetic catalog placement is unqualified.

Publication and source evidence

Checked source 1e05eff7c05b2e6007bfd70fa1559d6631974b0d and the merge both have tree 778d0135e48a210c40766710fec616f2fe08ba54. A PR OPEN field in the immutable package records its finalization-time status; the current merged state is separately evidenced above. Passing source CI establishes the code checks, not production model safety or real task completion.

Explicit use by people and agents

Use Go 1.27.1 and the merged source above for an explicitly selected local experiment. First verify the downloaded model's size and SHA-256.

hf download JooYoon/riidolaya-statehint-go-v0.3 statehint.rsh \
  --revision cf6deca9eb2c9919b1bbbbe82ce3576453d67a40 \
  --local-dir ./models/statehint-v0.3
go run ./cmd/riido-statehint --model ./models/statehint-v0.3/statehint.rsh \
  --text 'Please explain which part of the work this report covers.'

This explicitly selects a classification experiment. It does not change actual Riido labels or emoji, or establish task completion. Agents should retain the association between model version and source message and preserve uncertainty. These commands were not executed during this Wiki update; no new inference was performed.

Training reproduction instructions and the four data files are in the model card. The original plan preserves its relative cache paths; the separate public reproduction plan changes only four data paths to public filenames. The original driver is byte-identical to public cmd/riido-statehint-tune/main.go, SHA-256 1d9de51793bc3440978efd16a72d1d3eed6ff4c9785bf797a06bb21b67153d15. All 24 original source inputs retain plan SHA pins and public repository references.

Data, licensing and interpretation limits

Train 400 / validation 80 / calibration 80 / test 200 are 760 original synthetic messages by two AI authors. They are not actual user work, product ground truth, independent human annotations or statistically independent production samples. The test is now exposed development evidence and must not become the next candidate's fresh blind test.

Original synthetic text, annotations and the public Go code declare Apache-2.0. LICENSE and the original NOTICE are included. No private application source, actual work content, credentials, host paths or raw profiles were published.

The original macOS Go process's OS peak RSS was 15,974,400 bytes. Its scope includes all 4 fits, validation, calibration, test and model reload. It is not isolated model memory, Go heap or GPU usage. Reloaded-model probabilities matched across all 200 cases; that observation is separate from actual Riido integration and generalization evidence.