SearchingBinary/FastVLM-1.5B-Stage3-meeting-tiles-v5

FastVLM 1.5B - Meeting Tile v5 LoRA Adapter

LoRA adapter fine-tuned on FastVLM-1.5B-Stage3 for meeting screenshot analysis using a tile-crop strategy.

Tasks

• Platform detection: Identify meeting platform (Zoom, Teams, Google Meet, Webex) from full screenshots
• Tile name: Read participant name from an isolated tile crop
• Tile speaker: Detect active speaker (highlighted border) + read name from a tile crop

Training Details

• Base model: zhaode/FastVLM-1.5B-Stage3
• Method: LoRA (rank=8, alpha=16) on language model q/k/v/o projections
• Dataset: 103K synthetic examples (87.5K train / 15.5K val)
  • 34K full screenshots (platform detection)
  • 34K speaker tile crops (tile_speaker)
  • 34K non-speaker tile crops (tile_name)
  • 1K slide OCR examples
• Training: 2 epochs, batch_size=8, grad_accum=2, lr=2e-5, cosine schedule
• Hardware: NVIDIA A40 (46GB), ~4.5 hours
• Final loss: ~0.10-0.12
• Trainable params: 2.18M / 1.91B (0.11%)

Tile-Crop Strategy

Instead of asking the model to identify speakers from a full meeting screenshot (where name labels are ~13px after resize), we crop individual participant tiles and render them at full viewport (1280x720). This gives the model ~130px name labels -- a 10x improvement in readability.

Framework versions

• PEFT 0.18.1
• Transformers
• PyTorch