A Selector turns a per-sample or per-group score into a set of kept indices; dropped samples contribute zero gradient. In RL this removes a sample's contribution to the update after rollout.

weight 0 ≠ dropping rollout cost

Selecting here (post-advantage) removes the gradient but does not save the rollout — generation already happened. The rollout-saving variant lives at the replay-buffer / pre-rollout layer.

Enable via config

trainer: {v1: {trainer_mode: dataflex_sync}}
dataflex:
  mechanism: select
  scorer:   {name: group_solve_rate, params: {success_threshold: 0.5}}
  actuator: {name: threshold_band, params: {low: 0.0, high: 1.0}}

Run

python3 -m verl.trainer.main_ppo \
    algorithm.adv_estimator=grpo \
    ... (usual verl args) ... \
    trainer.v1.trainer_mode=dataflex_sync \
    +dataflex.mechanism=select \
    +dataflex.scorer.name=group_solve_rate \
    +dataflex.scorer.params.success_threshold=0.5 \
    +dataflex.actuator.name=threshold_band \
    +dataflex.actuator.params.low=0.0 \
    +dataflex.actuator.params.high=1.0

Runnable: examples/run_select_grpo.sh.

What to look for

dataflex/kept_frac:0.078...

the fraction of samples kept after filtering.

Available selectors

Name	Signal (scorer)	Rule
threshold_band	group_solve_rate	keep score ∈ (low, high)
topk_fraction	advantage_magnitude	keep top/bottom fraction
gfpo	reward_difficulty	per-group top-k by -L (short) or R/L (efficiency)
max_variance	reward_difficulty	per-group max-variance subset (PODS)

DAPO dynamic sampling = group_solve_rate + threshold_band(0,1) (drop all-solved / all-failed groups). Online Difficulty Filtering = the same with band (0.2, 0.8) — see its algorithm page.