JaneDing2025/simact_video_repro

SimAct-Video repro_gemini — model checkpoints (Fir H100)

NEPA-SimAct (nepa-large backbone) trained on the collaborator's Vision-Gemini annotations (SimAct-Video repro). seed 42, batch 32 x grad-accum 8 = 256 global, 10,000 steps, checkpoint every 1,000 steps. Commit 73ecfbd of Mars-tin/simact-dev, pinned nepa.sif (torch 2.8.0 / cu128).

Layout

<setting>/<mixture>/checkpoint-<step>/ — weights (model.safetensors) + config + tokenizer/vocab. Optimizer states are not included (inference / analysis only).

3 settings x 4 mixtures = 12 runs, 10 checkpoints each (steps 1000..10000):

Settings:

• cosine_completion — visual loss cosine, completion regime
• infonce_completion — visual loss infonce, completion regime
• infonce_fullseq — visual loss infonce, full-sequence regime

Mixtures (LM-forward / sim-loss weights):

• lmfwd100 — lmfwd 1.0 (no visual/sim loss)
• lmfwd50_simbwd50 — lmfwd 0.5, simbwd 0.5
• lmfwd50_simfwd50 — lmfwd 0.5, simfwd 0.5
• lmfwd50_simbwd25_simfwd25 — lmfwd 0.5, simbwd 0.25, simfwd 0.25