SII-liyang2024/jev-train

JEV decision and progress SFT v6

This dataset repository contains four configs for JEV supervised fine-tuning. The existing ahat and partnr configs are decision-choice data. This update adds ahat_progress and partnr_progress, which classify progress on a fixed 21-point grid (A through U, 0.00 through 1.00). The ahat_progress config uses the scene-prefix build v6_scene_prefix_choice21_20261027; it supersedes the previously uploaded absolute-prefix payload under the same config name.

config	purpose	train / validation / calibration / test rows
ahat	AHAT next-action decision	513,140 / 29,945 / 44,251 / 28,397
partnr	PARTNR next-action decision	5,585 / 583 / 279 / 193
ahat_progress	AHAT reference-plan progress choice	513,140 / 29,945 / 44,251 / 28,397
partnr_progress	PARTNR task-completion progress choice	23,121 / 2,842 / 1,614 / 1,318

Loading

from datasets import load_dataset

progress = load_dataset("SII-liyang2024/jev-train", "ahat_progress")
partnr_progress = load_dataset("SII-liyang2024/jev-train", "partnr_progress")
print(progress["train"][0]["messages"])

All four configs retain the custom calibration split. The progress trainer records contain messages, candidate_count (always 21), and target_index; the assistant answer is one option code from A through U.

Progress label semantics

• AHAT progress uses the scene-prefix input: the original HF scene_graph plus raw plan[:t] actions, with the reference-plan position t/T quantized to the nearest 0.05 with midpoint-up rounding. It is a weak proxy for reference-plan progress and does not certify natural-language task completion. The scene-prefix audit reports 615,733 samples and zero input/target conflicts; see metadata/ahat_progress/manifest.json and audit.json.
• PARTNR progress uses the original PARTNR environment task-completion fraction, quantized to the nearest 0.05 with midpoint-up rounding. It preserves semantic scenes, history, feedback, and splits; see metadata/partnr_progress/manifest.json and audit.json.

The raw continuous progress evidence in the source samples/ assets is not included in the public data/ configs. The uploaded trainer files contain the model input and categorical target only. Metadata examples and audits are retained for provenance.

Upload update

This directory is an update bundle for the existing repository. From this directory:

hf upload SII-liyang2024/jev-train . . --repo-type dataset --commit-message "Add AHAT and PARTNR progress SFT configs"

The large JSONL files are hard links to validated source files, so uploading this directory does not create another local copy. The existing decision files on the Hub remain in place; the README front matter lists all four configs.

Files

• data/ahat_progress/*.jsonl: AHAT scene-prefix progress-choice trainer records.
• data/partnr_progress/*.jsonl: PARTNR progress-choice trainer records.
• metadata/*/manifest.json, audit.json: source counts, labels, hashes, and audit results. metadata/ahat_progress/source_tasks.jsonl is source lineage only and is not a model input.
• metadata/*/candidates.json: fixed 21-option candidate grid.
• metadata/*/codebook.json, domain.pddl, token_encoding_audit.json: encoding and domain evidence where applicable.
• export_manifest.json: upload file sizes, row counts, and source SHA-256 values.

No model checkpoint or tokenizer is included. The progress targets are supervised categorical labels for candidate cross-entropy, not scalar-head regression targets.