rao254/manual-decisions-pilot

Manual decisions pilot

A pilot of typed decisions from selected prose in the supplied ClassNK freight-container and integrated-fire-control manuals. Exact counts and source hashes are in manifest.json. This is not a conversion of every section or a reproduction of LocalLLaMA/typed-decisions.

One row represents one decision. state and q are JSON strings so Hugging Face can store differing criteria shapes without changing their order. Decode both before passing them to build_training_item in training/train.py. target is an integer option index; type is choice, noul, or score. Choice criteria are alphabetically ordered; noul order is false, true; score options follow their list order. Metadata is for review, not model input.

Each scenario supplies a topic-choice row, a positive and negative evidence-support row, and a positive and negative evidence-relevance row. The relevance rubric has four levels, but this initial pilot contains only 0 and 3 labels. Topic catalogs are provisional and differ from the application's complete routing catalogs. Evidence-support nouls also differ from the application's evidence-needs nouls. This pilot tests domain decision training; it is not a drop-in replacement for the application's retrieval policy.

Labels and negative-pair choices are assistant-authored drafts requiring human review. Gold targets are hard labels, not teacher-sampled uncertainty. Sources use whitespace-normalized verbatim excerpts from machine-extracted prose. No table or drawing interpretation is included. Source pages, questions and scenario groups are separated across splits. Some topic classes occur only in validation; these small splits are not a class-balanced performance benchmark.

This dataset trains typed decisions with evidence provided for support/relevance judgments. It does not teach free-text answer generation or guarantee memorization of the manuals. Add diverse, reviewed cases, particularly partial evidence and all topic classes, before substantive training.

The source manuals are 530_freightcontainers_e_202606.pdf and 490_firecontrol_e_202606.pdf. Redistribution rights have not been established; no open license is asserted. Confirm permission for the included manual excerpts before public publication. No files are uploaded by the build scripts.

Load for the existing trainer

import json
from datasets import load_dataset

rows = load_dataset("YOUR_ACCOUNT/YOUR_DATASET", split="train")
sample_data = [
    {"state": json.loads(r["state"]), "q": json.loads(r["q"]),
     "target": r["target"], "type": r["type"]}
    for r in rows
]

Replace the hardcoded sample_data block with this loader. Keep validation separate. The current trainer batches the entire dataset at once: add minibatching before scaling. Token lengths must be checked with the exact training tokenizer and limits. The dataset build does not download a model or run GPU training.

Local validation of the initial pilot

All 80 rows passed the installed Laya render_options and build_sequence functions with the cached multilingual tokenizer at 512/192 limits; the longest sequence was 153 tokens and no state was truncated. The trainer's default checkpoint/tokenizer may differ, so repeat this check in the training environment. Hugging Face Datasets loading and GPU optimization were not executed locally.