Repository navigation
Assessment Pipeline: Add audio/video support #1184
Description
Is your feature request related to a problem?
The AI Assessments pipeline currently lacks support for audio and video modalities. This limits the assessment's effectiveness and prevents a comprehensive analysis of inputs.
Describe the solution you'd like
- Extend multimodal support to include audio/video in AI Assessments.
- Experiment with Gemini for video handling and explore methods with OpenAI/Anthropic.
- Sample an audio/video dataset from partners, run the pipeline, conduct human evaluations, and iterate.
- Support a language mix of ~50–60% English, ~10–15% Tamil, and a strong presence of South Indian languages like Telugu.
Original issue
Context
Extend multimodal support to include audio/video in AI Assessments.
Investigation
- Video: Gemini does appear to support video directly. Experiments to be done to figure out reasonable way of handling video in Open AI / Anthropic (sampling frames + full audio transcript ?)
Approach / acceptance criteria
- Sample audio/video dataset from partners, run pipeline, run human evals, iterate. Building the tech is easy; evaluation is the bottleneck.
- Language mix to support: ~50–60%+ English, Tamil ~10–15%, strong South Indian language presence, Telugu
Activity
[-]Add audio/video multimodal support to the AI Assessments pipeline[/-][+]Assessment Pipeline: Add audio/video support[/+]