WinterCCCCC/directorai-ref2va-lora

DirectorAI Ref2VA LoRA — full-3738w-8xh100

LoRA adapters for MiniMax-H3 Ref2VA, from the experiment full-3738w-8xh100. This repository holds intermediate checkpoints released for evaluation while training is still running; it is not a final model.

These are LoRA adapters only. They carry no base weights; using them requires the MiniMax-H3 Ref2VA base model, whose own terms govern that part.

Checkpoints

file	step	stage	validation loss
full-3738w-8xh100-step00000288.safetensors	288	end of the 8-card segment; the comparison baseline	0.462827
full-3738w-8xh100-step00000480.safetensors	480	end of epoch 1 (+12 steps)	0.453399
full-3738w-8xh100-step00000936.safetensors	936	end of epoch 2 (exact boundary)	—
full-3738w-8xh100-step00001056.safetensors	1056	latest validated point	0.445345

Total budget is 1404 steps = 3 epochs x 468. Epoch 1 ends at step 468, but checkpoints are written every 24 steps, so 468 itself was never saved; step 480 is the first save after epoch 1 completed and therefore covers all of epoch 1's data. Epoch 2's boundary, step 936, is a multiple of 24 and was saved exactly. The final step-1404 adapter will be added when training finishes.

Training setup

• 3738 training windows, 245 held out for validation (24 of them used in the loop)
• 1404 optimizer steps (3 epochs), global batch 8 = 4 cards x 2 gradient accumulation, batch 1 per device
• 4 x H100 80GB, about 177-179 s per optimizer step
• LoRA rank 128, alpha 128; AdamW, lr 2e-5, cosine with min lr, weight decay 0.01
• in-loop validation on 24 held-out windows every 96 steps
• trainer: musubi-tuner pinned at c4dc6b8 plus local patches (validation under block swap, elementwise RoPE, LR schedule rescale on a resume with another card count, save before validation)
• steps 1-288 ran on 8 cards, 289 onward on 4 cards with the global batch held at 8

Validation loss history

step	loss/validation
0	0.875663
96	0.513813
192	0.475836
288	0.462827
384	0.456176
480	0.453399
576	0.451272
672	0.448584
768	0.448102
864	0.447192
960	0.446129
1056	0.445345

Per-branch at step 384: video 0.171422, audio 0.284754.

Note on reading these numbers: this is a flow-matching objective, and the loss curve says little about perceptual quality. Past step 288 it moves by about 0.001 per 96 steps. Treat the table as a record, not as an evaluation; prefer a qualitative comparison between the checkpoints above.

Integrity

16f22bb3910d6446e099e8f0ff69c825086c75ec1f19c2708648e198cb9e1407  full-3738w-8xh100-step00000288.safetensors
997b7864752fc3083af7e441896d376ab56b7d5ccc396140d73a4da6bf49d588  full-3738w-8xh100-step00000480.safetensors
54fa3428478511117a47cac4550789cff2e51605d0c1c76aadedddde7a6b7e85  full-3738w-8xh100-step00000936.safetensors
02ffc55364be20d54d1b688c32ecf1f8e66e1f297bd47cf78eb1b06978671f3b  full-3738w-8xh100-step00001056.safetensors

Each file is 1192827824 bytes.