xtc doc

doc/setting/sampling.md

@@ -49,7 +49,7 @@ We use an LLM to infer the next token via random sampling.
     (presence_penalty 0.1)
    )
 
-   ; DRY (Don't Repeat Yourself) sampling.
+   ; "Don't Repeat Yourself".
    (dry
     (window_length 1200)
     (multiplier 0.8)
@@ -67,17 +67,25 @@ We use an LLM to infer the next token via random sampling.
 (language
  ((infer_via sampling)
   (adjust_thru (())
-   ; Top-K Sampling. 1 makes sampling deterministic.
+   ; Top-K. 1 makes sampling deterministic.
    (top_k 40)
-   ; Top-P Sampling. 0.0 is a no-op.
+   ; Top-P. 0.0 is a no-op.
    (top_p 0.9)
-   ; Locally Typical Sampling cutoff. 1.0 is a no-op.
+   ; Locally Typical cutoff. 1.0 is a no-op.
    (typical_p 0.9)
-   ; Min-P Sampling. 1.0 is a no-op.
+   ; Min-P. 1.0 is a no-op.
    ; Cut out tokens whose probability is less
    ; than 5% relative to the most probable token.
    (min_p 0.05)
 
+   ; "Exclude Top Choices".
+   (xtc
+    ; Probability threshold (default is 0.15).
+    (threshold 0.15)
+    ; Probability of performing the exclusion on this pass (defult is 1.0).
+    (probability 0.5)
+   )
+
    ; Temperature is not as necessary when tokens are expcluded.
    ; It can be >1.0 when Min-P has already filtered out the nonsensical tokens.
    (temperature 1.2)

src/language/inference_schema.hh

@@ -24,8 +24,8 @@ struct PenalizeWith {
 };
 
 struct Xtc {
-  float probability = 0.0f;
-  float threshold = 0.1f;
+  float threshold = 0.15f;
+  float probability = 1.0f;
 };
 
 struct AdjustViaKind {