Upload training/configs/dataset_quick_train.toml with huggingface_hub

training/configs/dataset_quick_train.toml

@@ -0,0 +1,43 @@
+# MiniMax H3 quick run — mega_curated 3.9k clips, i2va, video-only (no audio in sources)
+# Multi-resolution: three area buckets (mandatory for H3 — RoPE grid is area-normalized)
+# Small/medium resolutions per user directive (1024x576 dropped).
+# Sources normalized to 24 fps CFR, square pixels (SAR 1:1) in /workspace/datasets/mega_24
+
+[general]
+caption_extension = ".txt"
+batch_size = 1
+enable_bucket = true
+bucket_no_upscale = true
+
+# ~55% — low bucket (motion is cheap at low res)
+[[datasets]]
+video_directory = "/workspace/datasets/train_quick/v288"
+cache_directory = "/workspace/cache/quick/v288"
+resolution = [512, 288]
+h3_target_mode = "video"
+source_fps = 24.0
+frame_extraction = "full"
+max_frames = 124
+num_repeats = 1
+
+# ~28% — main bucket
+[[datasets]]
+video_directory = "/workspace/datasets/train_quick/v384"
+cache_directory = "/workspace/cache/quick/v384"
+resolution = [640, 384]
+h3_target_mode = "video"
+source_fps = 24.0
+frame_extraction = "full"
+max_frames = 124
+num_repeats = 1
+
+# ~17% — medium anchor (prevents low-density specialization)
+[[datasets]]
+video_directory = "/workspace/datasets/train_quick/v480"
+cache_directory = "/workspace/cache/quick/v480"
+resolution = [832, 480]
+h3_target_mode = "video"
+source_fps = 24.0
+frame_extraction = "full"
+max_frames = 124
+num_repeats = 1