Configuration

The main config file is vagen/configs/vagen_multiturn.yaml.

Trainer

trainer:
  skip_special_tokens_val: False
  skip_special_tokens_train: False
  replace_image_tokens_for_logging: True
  log_image:
    enable: True
    max_pending: 2
    png_compress_level: 0

Parameter	Type	Default	Description
skip_special_tokens_val	bool	False	Skip special tokens when logging validation outputs
skip_special_tokens_train	bool	False	Skip special tokens when logging training outputs
replace_image_tokens_for_logging	bool	True	Replace <image> tokens with placeholder for cleaner logs
log_image.enable	bool	True	Enable image logging to local folder
log_image.max_pending	int	2	Maximum pending image uploads
log_image.png_compress_level	int	0	PNG compression level (0=none, 9=max)

Filter

filter:
  name: reward_variance
  filter_kwargs: {}
  enable: False

Parameter	Type	Default	Description
name	str	reward_variance	Filter name (must be registered via @register_filter)
filter_kwargs	dict	{}	Arguments passed to filter function as **kwargs
enable	bool	False	Enable/disable filtering

See Custom Filter for creating your own filters.

Built-in Filter Options

reward_variance

filter:
  name: reward_variance
  filter_kwargs:
    topk: 0.2      # Keep top 20% groups by variance
    ddof: 0        # Degrees of freedom for variance calculation
  enable: True

reward_variance_top_p

filter:
  name: reward_variance_top_p
  filter_kwargs:
    top_p: 0.9     # Keep groups until 90% cumulative variance
    ddof: 0
  enable: True