| # Native-Bird image-text alignment (AfriCaption / Flickr8k) | |
| # All paths are relative to the project root. | |
| seed: 42 | |
| data: | |
| dir: data/processed # train/val/test parquet from scripts/04_split.py | |
| image_dir: data/raw/flickr8k/Flicker8k_Dataset | |
| langs: [eng, ibo, hau, yor] # captions used for training (each image contributes one pair per kept language) | |
| max_text_len: 64 | |
| subset: null # int -> use only first N images per split (smoke tests) | |
| image_encoder: | |
| # Any open_clip model or HF CLIP/SigLIP checkpoint. Frozen; features are cached to disk once. | |
| backend: hf # hf | open_clip | |
| name: google/siglip-base-patch16-224 | |
| cache_dir: data/features # cached image features: <cache_dir>/<safe_name>/<split>.pt | |
| text_encoder: | |
| # Pick the arm. Each arm is a sentence-transformers model id. | |
| arm: native_bird # native_bird | embeddinggemma2 | embeddinggemma1 | labse | custom | |
| arms: | |
| native_bird: {name: Modularcomputing/Native-Bird, prompt: ""} | |
| embeddinggemma2: {name: google/embeddinggemma-2, prompt: "task: search result | query: ", text_only: true} | |
| embeddinggemma1: {name: google/embeddinggemma-300m, prompt: "task: search result | query: "} | |
| labse: {name: sentence-transformers/LaBSE, prompt: ""} | |
| custom: {name: null, prompt: ""} | |
| unfreeze_last_n_layers: 0 # 0 = fully frozen (text embeddings are cached); >0 fine-tunes the top-N transformer layers | |
| cache_dir: data/features | |
| model: | |
| proj_dim: 512 | |
| proj_hidden: 0 # 0 = linear projection; >0 = 2-layer MLP with this hidden size | |
| dropout: 0.1 | |
| init_logit_scale: 2.659 # ln(1/0.07), CLIP default; learnable, clamped at ln(100) | |
| train: | |
| epochs: 20 | |
| batch_size: 256 # contrastive batch; bigger = more negatives | |
| lr: 1.0e-3 # projection heads | |
| text_lr: 1.0e-5 # unfrozen text layers (only if unfreeze_last_n_layers > 0) | |
| weight_decay: 0.01 | |
| warmup_steps: 100 | |
| same_image_negatives: mask # mask: captions of the same image in other languages are not treated as negatives | |
| device: auto # auto | cuda | cpu | |
| num_workers: 4 | |
| out_dir: runs/${text_encoder.arm} | |
| eval_every: 1 | |
| eval: | |
| split: test | |
| ks: [1, 5, 10] | |
| langs: [eng, ibo, hau, yor] | |