Update README.md

README.md

@@ -37,41 +37,6 @@ This version is best when you want broad embeddings capturing item relationships
 
 ---
 
-## How to enable each version in code
-
-The behavior is controlled by the masking probabilities used in the collator:
-
-- `TRAIN_RANDOM_MLM_PROB`
-- `EVAL_RANDOM_MLM_PROB`
-
-And by whether you “force-mask last event items” (enabled in the forecasting collator logic).
-
-### Recommended settings
-
-#### Forecasting-only (Version A)
-- Train: `TRAIN_RANDOM_MLM_PROB = 0.0` (no random MLM noise)
-- Eval:  `EVAL_RANDOM_MLM_PROB  = 0.0`
-- Force-masking last event `ITEM_*` stays **ON**
-
-This focuses learning and evaluation on last-event item prediction.
-
-#### Forecasting + regularization (Version A + random noise)
-- Train: `TRAIN_RANDOM_MLM_PROB = 0.15`
-- Eval:  `EVAL_RANDOM_MLM_PROB  = 0.0`
-- Force-masking last event `ITEM_*` stays **ON**
-
-This is the default “forecasting twist” setup: train with extra random MLM, evaluate cleanly on forecasting.
-
-#### Random MLM across full sequence (Version B)
-- Train: `TRAIN_RANDOM_MLM_PROB = 0.15`
-- Eval:  `EVAL_RANDOM_MLM_PROB  = 0.15` (or any non-zero)
-- (Optional) disable force-masking last-event items if you want *pure* standard MLM
-
-> Note: In the current `ForecastingCollator`, force-masking last-event items is always applied.  
-> If you want **pure random MLM** (no forecasting), add a flag like `force_last_event=False` and skip the `prob[force_mask] = 1.0` step.
-
----
-
 ## What the forecasting masking means (in practice)
 
 A packed firm sequence looks like:
@@ -109,7 +74,6 @@ For forecasting, **Item@K** is the main metric because it directly measures how
 - **Acc@5:** `0.6651`
 - **Acc@10:** `0.6944`
 
-> “HARD” refers to the stricter evaluation setting used in our validation protocol (forecasting-focused metrics on masked targets).
 
 ---