adipotnis/trajreasoner-data

TrajReasoner training data

One repository contains the full enriched TerraSentia, GrandTour, and SCAND sources, the reasoning trace files, and the exact v8 SFT train/validation rows. Source configurations keep their native schemas, images, and baked labels. The default sft configuration has embedded images, prompt, and completion. REASONING/SOLUTION labels are preserved; the trainer renders native thinking.

from datasets import load_dataset
train = load_dataset("adipotnis/trajreasoner-data", "sft", split="train", revision="<verified commit>")
source = load_dataset("adipotnis/trajreasoner-data", "terrasentia", split="train", revision="<verified commit>")

Use scripts/build_dataset.py in TrajReasoner for all future builds. recipe.yaml records the mixture. manifest.json records pinned original source revisions, file hashes, row counts, and export verification. Original dataset cards are preserved under provenance/; their licenses and access conditions continue to apply. Dataset access requires manual approval.

The preserved v8 holdout is the historical per-loader split, not a new trajectory-disjoint benchmark. Legacy reasoning keys are positional; their original files are preserved rather than regenerated during migration.