utils.collators.mm_rl

utils.collators.mm_rl

On-the-fly image loading for TRL’s vision RL paths (DPO/IPO, KTO, GRPO).

TRL expects an images list of loaded images plus bare {"type": "image"} prompt placeholders; path/URL/base64 references are resolved here at collate time.

Classes

Name Description
AxolotlVisionPreferenceCollator TRL’s DPO vision collator with axolotl image loading/resizing.
AxolotlVisionUnpairedPreferenceCollator TRL’s KTO vision collator with axolotl image loading/resizing.
MultimodalRLExampleNormalizer Resolve image refs into loaded, optionally resized, images in TRL’s format.

AxolotlVisionPreferenceCollator

utils.collators.mm_rl.AxolotlVisionPreferenceCollator(
    normalizer=MultimodalRLExampleNormalizer(),
)

TRL’s DPO vision collator with axolotl image loading/resizing.

AxolotlVisionUnpairedPreferenceCollator

utils.collators.mm_rl.AxolotlVisionUnpairedPreferenceCollator(
    normalizer=MultimodalRLExampleNormalizer(),
)

TRL’s KTO vision collator with axolotl image loading/resizing.

MultimodalRLExampleNormalizer

utils.collators.mm_rl.MultimodalRLExampleNormalizer(
    image_size=None,
    image_resize_algorithm=None,
)

Resolve image refs into loaded, optionally resized, images in TRL’s format.