src/chug/doc/doc_read_processor.py (7 lines):
	- line 49: # FIXME for initial behaviour we will randomly sample one of N pages
	- line 50: # TODO determine if we want to train in multi-page mode, use another sampling strategy?
	- line 54: n_wanted_pages = min(1, num_pages)  # TODO increase for multi-page processing, rand start+end?
	- line 79: # FIXME currently encoding each page separately with own start/end tokens.
	- line 123: # FIXME cleanup, split process and decode for more flexibility
	- line 136: # FIXME warn assert that target not supported w/o text preprocessing?
	- line 172: # TODO support 'image info' to relay details such as text bbox, layout


src/chug/image/build_transforms_image.py (4 lines):
	- line 49: # FIXME map aug_cfg -> timm.create_transform args more carefully
	- line 54: use_prefetcher=not do_normalize,  # FIXME prefetcher mode disables normalize, but outputs np.array
	- line 75: use_prefetcher=not do_normalize,  # FIXME prefetcher mode disables normalize, but outputs np.array
	- line 122: # FIXME note we don't have good option for 'longest' resizing w/ RRC


src/chug/image/build_transforms_doc.py (4 lines):
	- line 71: # FIXME merge defaults w/ aug_cfg
	- line 205: # FIXME merge defaults w/ aug_cfg
	- line 289: alpha_affine=12.,  # FIXME no common param, alpha_affine unique to alb
	- line 324: # FIXME leave alb uncomposed too if composed=False?


src/chug/wds/decode.py (4 lines):
	- line 59: # FIXME test and use an alternate pdf reader/render as default
	- line 205: # FIXME need to work out padding / selection issues for multi-page support
	- line 208: assert page_sampling in {'random', 'first', 'last'}  # TODO add 'all' w/ multi-page support
	- line 298: enable_doc: bool = False,  # FIXME enable doc support by default once tested?


src/chug/hfds/loader.py (3 lines):
	- line 107: # FIXME num_samples calc, get a reliable estimate from dataset in streaming mode
	- line 120: # FIXME split_dataset_by_node has some concerns as currently implemented
	- line 133: drop_last=batched and is_training,  # FIXME improve wrt train vs validation vs sharding specifics


src/chug/doc/doc_vqa_processor.py (3 lines):
	- line 56: # FIXME support flexible q/a prompting formats, do with prefix/suffix or template strings?
	- line 68: # FIXME multi qa expansion is a WIP
	- line 108: # FIXME implement expansion for multi-qa (and eventually multi-page option)


src/chug/loader.py (2 lines):
	- line 72: # FIXME make validation task specific once we have tasks that don't require both image and text preproc
	- line 121: multi_interval=True, # FIXME via config?


src/chug/common/config.py (2 lines):
	- line 227: # FIXME add code to resolve shard information from _info.yaml or .json files (see dataset_info.py)
	- line 255: sizes = None  # FIXME resolve sizes


src/chug/doc/doc_processor.py (2 lines):
	- line 105: # FIXME note, should move to torchvision v2 annotations at some point
	- line 192: # FIXME separate decode & preprocess interfaces


src/chug/text/tokenization.py (2 lines):
	- line 92: # FIXME calculate prompt end pos for validation use (target not needed)
	- line 114: # FIXME just binding prepare_text_input fn for now


src/chug/wds/shardlists.py (1 line):
	- line 26: self.num_sub_intervals = num_sub_intervals  # FIXME experimental feature


src/chug/wds/dataset_info.py (1 line):
	- line 30: ## FIXME this is not working / not completed, parsing _info files is a TODO


src/chug/task_pipeline/pipeline_image_text.py (1 line):
	- line 40: # FIXME add support for caption target for caption tasks or use a separate pipe?


src/chug/common/types.py (1 line):
	- line 67: # TODO resolve dataset info and track base url, shard info (sizes, etc)


src/chug/task_pipeline/pipeline_gtparse.py (1 line):
	- line 39: FIXME add support for caption target for caption tasks or separate pipe?


src/chug/wds/tariterators.py (1 line):
	- line 49: # FIXME wds version throws if suffix in current_sample, but we have a potential for


src/chug/image/transforms_torch.py (1 line):
	- line 116: # FIXME check tensor vs PIL and convert as needed, this is assuming PIL right now


src/chug/wds/loader.py (1 line):
	- line 104: # FIXME support distributed eval


src/chug/task_pipeline/pipeline_doc_vqa.py (1 line):
	- line 74: # FIXME prompt templates instead of prefix+suffix above?