trainer corpus suite

Small specialized models are only as good as their corpora, and quality multimodal training data is the hardest asset to source. Public datasets rarely match the target task, and frontier-model annotation is expensive and unreliable in ways that repeat across models. The suite exists to turn raw video into structured, verified training data; every annotation is cross-checked against independently extracted signal.

• Scene detection and segmentation — long-form video decomposed into coherent clips with shot-boundary awareness.
• Parallel signal extraction — independent analyzers run per clip: pose estimation, action recognition, emotion recognition over time, object detection, and domain-specific measurement (the biomechanics pipeline plugs in here). Each produces a structured, timestamped signal track.
• VLM synthesis — vision-language models generate reasoning-rich annotations conditioned on the signal tracks as well as the pixels, grounding the language in independently measured evidence.
• Cross-validation and review — multi-pass generation, cross-model critique, and quality classification decide what enters the corpus; disagreement between signals and captions is surfaced for review.
• Corpus assembly — structured chain-of-thought training rows, length-audited and schema-validated for supervised fine-tuning of small VLMs.

The suite's failure analysis of VLM captioning is what defined the problem space that became REVEAL: models confabulate motion, miss the subject, or narrate from priors instead of pixels. The cross-validation patterns built here, independent signal tracks anchoring generated language, are the prompt-level ancestor of REVEAL's exogenous conditioning.

Portions of the suite are being prepared for public release; the repository below reflects the currently-public subset.

→ github.com/aiaxmaior/trainer_corpus_suite