Model card for mammut2_ViT-B-32.cc12m

MaMMUT2 B/32: image encoder + a single text decoder run twice per step (text-only contrastive pass, then image-cross-attention caption pass).

Research checkpoint from a B/32-scale architecture/objective comparison on CC12M (CLIP vs MaMMUT2 vs modern text decoder); not a production model.

Model Details

• Dataset: CC12M
• Epochs: 32
• Precision: bf16 amp
• Weight decay: 0.25
• Objective: InfoNCE + 2x caption CE + 1e-5 z-loss
• Text decoder: classic transformer, cross-attn every 2nd block, BPE vocab 49408, ctx 77
• Optimizer: adamw, lr 3e-3
• Batch size: 1024

Model Usage

import torch
import open_clip

model, preprocess = open_clip.create_model_from_pretrained('hf-hub:rwightman/mammut2_ViT-B-32.cc12m')
tokenizer = open_clip.get_tokenizer('hf-hub:rwightman/mammut2_ViT-B-32.cc12m')

image = preprocess(pil_image).unsqueeze(0)
text = tokenizer(['a dog', 'a cat', 'a bird'])
with torch.no_grad():
    image_features = model.encode_image(image)
    text_features = model.encode_text(text)

Caption generation:

with torch.no_grad():
    tokens = model.generate(image, generation_type='beam_search', seq_len=30)
print(tokenizer.decode(tokens[0].cpu().tolist()))

Model Comparison

run	IN1K	IN-V2	IN-R	IN-Sk	I2T R@1	T2I R@1	CIDEr	BLEU-4	METEOR
clip-siglip	42.72	36.44	54.38	32.08	34.84	25.95
clip-naflex	40.4	34.43	50.53	28.55	34.6	24.25
mammut2 (no z)	37.8	31.98	48.8	27.16	37.52	24.72	10.02	2.43	7.37
mammut2 (+z) (this model)	38.36	32.47	49.44	27.12	38.18	24.54	10.04	2.41	7.38
mammut2-naflex	37.54	31.92	48.63	26.28	38.64	24.7	9.52	2.33	7.4
mammut2-mt no-qk lr1e-3	38.07	32.78	50.84	27.28	37.44	25.18	10.96	3.58	7.69
mammut2-mt qk lr2e-3	38.37	33.02	50.56	27.18	37.94	24.93	10.85	3.44	7.57
mammut2-mt qk lr3e-3	37.05	31.82	48.41	26.03	35.92	24.06	10.29	3.31	7.35

All eight runs: CC12M, 32 epochs, 2x GPU, bf16 amp, wd 0.25, evaluated at epoch 32 with one harness (ImageNet zero-shot w/ OpenAI 80-prompt ensemble; COCO Karpathy-test retrieval, I2T = any-of-5-captions; caption metrics vs the 5 COCO refs, beam search). Key reads:

• The CLIP rows are prior runs, not purpose-trained controls: they were the closest available baselines but come from NAdaMuon optimizer experiments (nadamuon @ bs 4096 vs the MaMMUT runs' adamw @ bs ~1024), so cross-family deltas bundle optimizer + batch + loss.
• Caption objective -> I2T retrieval: every MaMMUT2 run beats both CLIP runs by 1-4 pt I2T R@1 despite 1/4 the contrastive batch and a plainer optimizer.
• Zero-shot ordering tracks batch/loss/optimizer, not architecture.
• NaFlex vision and 1e-5 caption z-loss are both ~free (within the <=0.6 pt repeat-run noise).
• Modern text tower (bundled with r50k tokenizer + ctx 128) at its best LR: IN1K parity with classic, +V2/+IN-R/+T2I, and clearly better caption metrics. qk-norm buys 2x LR headroom rather than direct quality; 3e-3 is past the stable edge even with it.
• Caption scores are style-limited: CC12M alt-text register vs COCO references. Absolute CIDEr is low for all runs; relative differences are the signal.