Add 510-fable_more-shorter-rounds_s1
Browse files- runs/510-fable_more-shorter-rounds_s1/eval_dev.jsonl +0 -0
- runs/510-fable_more-shorter-rounds_s1/final.pt +3 -0
- runs/510-fable_more-shorter-rounds_s1/found_targets.jsonl +0 -0
- runs/510-fable_more-shorter-rounds_s1/lean_gate.disagree.jsonl +19 -0
- runs/510-fable_more-shorter-rounds_s1/lean_gate.jsonl +60 -0
- runs/510-fable_more-shorter-rounds_s1/pretrain.pt +3 -0
- runs/510-fable_more-shorter-rounds_s1/source/harness.py +529 -0
- runs/510-fable_more-shorter-rounds_s1/source/pretrain.py +611 -0
- runs/510-fable_more-shorter-rounds_s1/source/remote.sh +29 -0
- runs/510-fable_more-shorter-rounds_s1/summary.json +2321 -0
runs/510-fable_more-shorter-rounds_s1/eval_dev.jsonl
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|
runs/510-fable_more-shorter-rounds_s1/final.pt
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:c11a64164aaa98001e90c2b584f16d22f483d509c89a33a1fdf6036a9daa827e
|
| 3 |
+
size 25624925
|
runs/510-fable_more-shorter-rounds_s1/found_targets.jsonl
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|
runs/510-fable_more-shorter-rounds_s1/lean_gate.disagree.jsonl
ADDED
|
@@ -0,0 +1,19 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{"prompt": "THM Q , ( Q > F ) SEQ ( ( ~ Q ) & Q ) PRF", "lean_text": "have n25 : Q := h1 ; have n26 : ( Q → False ) := h2 ; have n27 : ( ¬ Q ) := n26 ; have n28 : Q := n25 ; have n29 : ( ( ¬ Q ) ∧ Q ) := ⟨ n27 , n28 ⟩ ; exact n29", "nd": "N1 Q : PR ; N2 ( Q > F ) : PR ; N3 ( ~ Q ) : R N2 ; N4 Q : R N1 ; N5 ( ( ~ Q ) & Q ) : ANDI N3 N4 ; QED", "nd_ok": false, "lean_ok": true}
|
| 2 |
+
{"prompt": "THM ( P > ( ( Q > R ) v ( Q & P ) ) ) SEQ ( ( ~ ( ( Q > R ) v ( Q & P ) ) ) > ( ~ P ) ) PRF", "lean_text": "have n54 : ( P → ( ( Q → R ) ∨ ( Q ∧ P ) ) ) := h1 ; have n55 : ( ( ¬ ( ( Q → R ) ∨ ( Q ∧ P ) ) ) → ( ¬ P ) ) := ( fun ( n56 : ( ¬ ( ( Q → R ) ∨ ( Q ∧ P ) ) ) ) => by have n57 : ( ¬ P ) := ( fun ( n58 : P ) => by have n59 : ( ( Q → R ) ∨ ( Q ∧ P ) ) := n54 n58 ; have n60 : False := n56 n59 ; exact n60 ) ; exact n57 ) ; exact n55", "nd": "N1 ( P > ( ( Q > R ) v ( Q & P ) ) ) : PR ; N2 | ( ~ ( ( Q > R ) v ( Q & P ) ) ) : AS ; N3 | | P : AS ; N4 | | ( ( Q > R ) v ( Q & P ) ) : IMPE N1 N3 ; N5 | | F : NEGE N4 N2 ; N6 | ( ~ P ) : IMPI N3 N5 ; N7 ( ( ~ ( ( Q > R ) v ( Q & P ) ) ) > ( ~ P ) ) : IMPI N2 N6 ; QED", "nd_ok": false, "lean_ok": true}
|
| 3 |
+
{"prompt": "THM ( ~ ( Q > ( P v Q ) ) ) , ( S v P ) SEQ ( ( ~ S ) v ( ~ R ) ) PRF", "lean_text": "have n10 : ( ¬ ( Q → ( P ∨ Q ) ) ) := h1 ; have n11 : ( S ∨ P ) := h2 ; have n12 : ( ¬ R ) := ( fun ( n13 : R ) => by have n14 : ( Q → ( P ∨ Q ) ) := ( fun ( n15 : Q ) => by have n16 : ( P ∨ Q ) := Or.inr n15 ; exact n16 ) ; have n17 : False := n10 n14 ; exact n17 ) ; have n18 : ( ( ¬ S ) ∨ ( ¬ R ) ) := Or.inr n12 ; exact n18", "nd": "N1 ( ~ ( Q > ( P v Q ) ) ) : PR ; N2 ( S v P ) : PR ; N3 | R : AS ; N4 | | Q : AS ; N5 | | ( P v Q ) : ORI2 N4 ; N6 | ( Q > ( P v Q ) ) : IMPI N4 N5 ; N7 | F : NEGE N6 N1 ; N8 ( ~ R ) : IMPI N3 N7 ; N9 ( ( ~ S ) v ( ~ R ) ) : ORI2 N8 ; QED", "nd_ok": false, "lean_ok": true}
|
| 4 |
+
{"prompt": "THM ( P > ( ( S v Q ) v ( Q & P ) ) ) SEQ ( ( ~ ( ( S v Q ) v ( Q & P ) ) ) > ( ~ P ) ) PRF", "lean_text": "have n54 : ( P → ( ( S ∨ Q ) ∨ ( Q ∧ P ) ) ) := h1 ; have n55 : ( ( ¬ ( ( S ∨ Q ) ∨ ( Q ∧ P ) ) ) → ( ¬ P ) ) := ( fun ( n56 : ( ¬ ( ( S ∨ Q ) ∨ ( Q ∧ P ) ) ) ) => by have n57 : ( ¬ P ) := ( fun ( n58 : P ) => by have n59 : ( ( S ∨ Q ) ∨ ( Q ∧ P ) ) := n54 n58 ; have n60 : False := n56 n59 ; exact n60 ) ; exact n57 ) ; exact n55", "nd": "N1 ( P > ( ( S v Q ) v ( Q & P ) ) ) : PR ; N2 | ( ~ ( ( S v Q ) v ( Q & P ) ) ) : AS ; N3 | | P : AS ; N4 | | ( ( S v Q ) v ( Q & P ) ) : IMPE N1 N3 ; N5 | | F : NEGE N4 N2 ; N6 | ( ~ P ) : IMPI N3 N5 ; N7 ( ( ~ ( ( S v Q ) v ( Q & P ) ) ) > ( ~ P ) ) : IMPI N2 N6 ; QED", "nd_ok": false, "lean_ok": true}
|
| 5 |
+
{"prompt": "THM ( P > ( ( S v Q ) v ( Q & P ) ) ) SEQ ( ( ~ ( ( S v Q ) v ( Q & P ) ) ) > ( ~ P ) ) PRF", "lean_text": "have n38 : ( P → ( ( S ∨ Q ) ∨ ( Q ∧ P ) ) ) := h1 ; have n39 : ( ( ¬ ( ( S ∨ Q ) ∨ ( Q ∧ P ) ) ) → ( ¬ P ) ) := ( fun ( n40 : ( ¬ ( ( S ∨ Q ) ∨ ( Q ∧ P ) ) ) ) => by have n41 : ( ¬ P ) := ( fun ( n42 : P ) => by have n43 : ( ( S ∨ Q ) ∨ ( Q ∧ P ) ) := n38 n42 ; have n44 : False := n40 n43 ; exact n44 ) ; exact n41 ) ; exact n39", "nd": "N1 ( P > ( ( S v Q ) v ( Q & P ) ) ) : PR ; N2 | ( ~ ( ( S v Q ) v ( Q & P ) ) ) : AS ; N3 | | P : AS ; N4 | | ( ( S v Q ) v ( Q & P ) ) : IMPE N1 N3 ; N5 | | F : NEGE N4 N2 ; N6 | ( ~ P ) : IMPI N3 N5 ; N7 ( ( ~ ( ( S v Q ) v ( Q & P ) ) ) > ( ~ P ) ) : IMPI N2 N6 ; QED", "nd_ok": false, "lean_ok": true}
|
| 6 |
+
{"prompt": "THM ( ~ ( ( ~ ( ~ ( R & Q ) ) ) > ( ( ~ ( R & Q ) ) > ( ( ~ ( R & Q ) ) v ( ( Q > R ) & ( P > R ) ) ) ) ) ) , ( ( ~ ( R & S ) ) v P ) SEQ ( ( P & P ) & ( R v P ) ) PRF", "lean_text": "have n24 : ( ¬ ( ( ¬ ( ¬ ( R ∧ Q ) ) ) → ( ( ¬ ( R ∧ Q ) ) → ( ( ¬ ( R ∧ Q ) ) ∨ ( ( Q → R ) ∧ ( P → R ) ) ) ) ) ) := h1 ; have n25 : ( ( ¬ ( ¬ ( R ∧ Q ) ) ) → ( ( ¬ ( R ∧ Q ) ) → ( ( ¬ ( R ∧ Q ) ) ∨ ( ( Q → R ) ∧ ( P → R ) ) ) ) ) := ( fun ( n26 : ( ¬ ( ¬ ( R ∧ Q ) ) ) ) => by have n27 : ( ( ¬ ( R ∧ Q ) ) → ( ( ¬ ( R ∧ Q ) ) ∨ ( ( Q → R ) ∧ ( P → R ) ) ) ) := ( fun ( n28 : ( ¬ ( R ∧ Q ) ) ) => by have n29 : ( ( ¬ ( R ∧ Q ) ) ∨ ( ( Q → R ) ∧ ( P → R ) ) ) := Or.inl n28 ; exact n29 ) ; exact n27 ) ; have n30 : False := n24 n25 ; have n31 : ( ( P ∧ P ) ∧ ( R ∨ P ) ) := False.elim n30 ; exact n31", "nd": "N1 ( ~ ( ( ~ ( ~ ( R & Q ) ) ) > ( ( ~ ( R & Q ) ) > ( ( ~ ( R & Q ) ) v ( ( Q > R ) & ( P > R ) ) ) ) ) ) : PR ; N2 | ( ~ ( ~ ( R & Q ) ) ) : AS ; N3 | | ( ~ ( R & Q ) ) : AS ; N4 | | ( ( ~ ( R & Q ) ) v ( ( Q > R ) & ( P > R ) ) ) : ORI1 N3 ; N5 | ( ( ~ ( R & Q ) ) > ( ( ~ ( R & Q ) ) v ( ( Q > R ) & ( P > R ) ) ) ) : IMPI N3 N4 ; N6 ( ( ~ ( ~ ( R & Q ) ) ) > ( ( ~ ( R & Q ) ) > ( ( ~ ( R & Q ) ) v ( ( Q > R ) & ( P > R ) ) ) ) ) : IMPI N2 N5 ; N7 F : NEGE N6 N1 ; N8 ( ( P & P ) & ( R v P ) ) : BOTE N7 ; QED", "nd_ok": false, "lean_ok": true}
|
| 7 |
+
{"prompt": "THM ( ~ ( ( ~ ( ~ ( R & Q ) ) ) > ( ( ~ ( R & Q ) ) > ( ( ~ ( R & Q ) ) v ( ( Q > R ) & ( P > R ) ) ) ) ) ) , ( ( ~ ( R & S ) ) v P ) SEQ ( ( P & P ) & ( R v P ) ) PRF", "lean_text": "have n22 : ( ¬ ( ( ¬ ( ¬ ( R ∧ Q ) ) ) → ( ( ¬ ( R ∧ Q ) ) → ( ( ¬ ( R ∧ Q ) ) ∨ ( ( Q → R ) ∧ ( P → R ) ) ) ) ) ) := h1 ; have n23 : ( ( ¬ ( ¬ ( R ∧ Q ) ) ) → ( ( ¬ ( R ∧ Q ) ) → ( ( ¬ ( R ∧ Q ) ) ∨ ( ( Q → R ) ∧ ( P → R ) ) ) ) ) := ( fun ( n24 : ( ¬ ( ¬ ( R ∧ Q ) ) ) ) => by have n25 : ( ( ¬ ( R ∧ Q ) ) → ( ( ¬ ( R ∧ Q ) ) ∨ ( ( Q → R ) ∧ ( P → R ) ) ) ) := ( fun ( n26 : ( ¬ ( R ∧ Q ) ) ) => by have n27 : ( ( ¬ ( R ∧ Q ) ) ∨ ( ( Q → R ) ∧ ( P → R ) ) ) := Or.inl n26 ; exact n27 ) ; exact n25 ) ; have n28 : False := n22 n23 ; have n29 : ( ( P ∧ P ) ∧ ( R ∨ P ) ) := False.elim n28 ; exact n29", "nd": "N1 ( ~ ( ( ~ ( ~ ( R & Q ) ) ) > ( ( ~ ( R & Q ) ) > ( ( ~ ( R & Q ) ) v ( ( Q > R ) & ( P > R ) ) ) ) ) ) : PR ; N2 | ( ~ ( ~ ( R & Q ) ) ) : AS ; N3 | | ( ~ ( R & Q ) ) : AS ; N4 | | ( ( ~ ( R & Q ) ) v ( ( Q > R ) & ( P > R ) ) ) : ORI1 N3 ; N5 | ( ( ~ ( R & Q ) ) > ( ( ~ ( R & Q ) ) v ( ( Q > R ) & ( P > R ) ) ) ) : IMPI N3 N4 ; N6 ( ( ~ ( ~ ( R & Q ) ) ) > ( ( ~ ( R & Q ) ) > ( ( ~ ( R & Q ) ) v ( ( Q > R ) & ( P > R ) ) ) ) ) : IMPI N2 N5 ; N7 F : NEGE N6 N1 ; N8 ( ( P & P ) & ( R v P ) ) : BOTE N7 ; QED", "nd_ok": false, "lean_ok": true}
|
| 8 |
+
{"prompt": "THM ( ( R > R ) > R ) , ( R > R ) SEQ ( ( P v ( Q > P ) ) > ( ( ~ P ) > ( ( Q v P ) > ( R & ( ( R > R ) > R ) ) ) ) ) PRF", "lean_text": "have n15 : ( ( R → R ) → R ) := h1 ; have n16 : ( R → R ) := ( fun ( n17 : R ) => by exact n17 ) ; have n18 : ( ( P ∨ ( Q → P ) ) → ( ( ¬ P ) → ( ( Q ∨ P ) → ( R ∧ ( ( R → R ) → R ) ) ) ) ) := ( fun ( n19 : ( P ∨ ( Q → P ) ) ) => by have n20 : ( ( ¬ P ) → ( ( Q ∨ P ) → ( R ∧ ( ( R → R ) → R ) ) ) ) := ( fun ( n21 : ( ¬ P ) ) => by have n22 : ( ( Q ∨ P ) → ( R ∧ ( ( R → R ) → R ) ) ) := ( fun ( n23 : ( Q ∨ P ) ) => by have n24 : R := n15 n16 ; have n25 : ( R ∧ ( ( R → R ) → R ) ) := ⟨ n24 , n15 ⟩ ; exact n25 ) ; exact n22 ) ; exact n20 ) ; exact n18", "nd": "N1 ( ( R > R ) > R ) : PR ; N2 | R : AS ; N3 ( R > R ) : IMPI N2 N2 ; N4 | ( P v ( Q > P ) ) : AS ; N5 | | ( ~ P ) : AS ; N6 | | | ( Q v P ) : AS ; N7 | | | R : IMPE N1 N3 ; N8 | | | ( R & ( ( R > R ) > R ) ) : ANDI N7 N1 ; N9 | | ( ( Q v P ) > ( R & ( ( R > R ) > R ) ) ) : IMPI N6 N8 ; N10 | ( ( ~ P ) > ( ( Q v P ) > ( R & ( ( R > R ) > R ) ) ) ) : IMPI N5 N9 ; N11 ( ( P v ( Q > P ) ) > ( ( ~ P ) > ( ( Q v P ) > ( R & ( ( R > R ) > R ) ) ) ) ) : IMPI N4 N10 ; QED", "nd_ok": false, "lean_ok": true}
|
| 9 |
+
{"prompt": "THM ( ( R > R ) > R ) , ( R > R ) SEQ ( ( P v ( Q > P ) ) > ( ( ~ P ) > ( ( Q v P ) > ( R & ( ( R > R ) > R ) ) ) ) ) PRF", "lean_text": "have n26 : ( ( R → R ) → R ) := h1 ; have n27 : ( R → R ) := ( fun ( n28 : R ) => by exact n28 ) ; have n29 : ( ( P ∨ ( Q → P ) ) → ( ( ¬ P ) → ( ( Q ∨ P ) → ( R ∧ ( ( R → R ) → R ) ) ) ) ) := ( fun ( n30 : ( P ∨ ( Q → P ) ) ) => by have n31 : ( ( ¬ P ) → ( ( Q ∨ P ) → ( R ∧ ( ( R → R ) → R ) ) ) ) := ( fun ( n32 : ( ¬ P ) ) => by have n33 : R := n26 n27 ; have n34 : ( ( Q ∨ P ) → ( R ∧ ( ( R → R ) → R ) ) ) := ( fun ( n35 : ( Q ∨ P ) ) => by have n36 : ( R ∧ ( ( R → R ) → R ) ) := ⟨ n33 , n26 ⟩ ; exact n36 ) ; exact n34 ) ; exact n31 ) ; exact n29", "nd": "N1 ( ( R > R ) > R ) : PR ; N2 | R : AS ; N3 ( R > R ) : IMPI N2 N2 ; N4 | ( P v ( Q > P ) ) : AS ; N5 | | ( ~ P ) : AS ; N6 | | R : IMPE N1 N3 ; N7 | | | ( Q v P ) : AS ; N8 | | | ( R & ( ( R > R ) > R ) ) : ANDI N6 N1 ; N9 | | ( ( Q v P ) > ( R & ( ( R > R ) > R ) ) ) : IMPI N7 N8 ; N10 | ( ( ~ P ) > ( ( Q v P ) > ( R & ( ( R > R ) > R ) ) ) ) : IMPI N5 N9 ; N11 ( ( P v ( Q > P ) ) > ( ( ~ P ) > ( ( Q v P ) > ( R & ( ( R > R ) > R ) ) ) ) ) : IMPI N4 N10 ; QED", "nd_ok": false, "lean_ok": true}
|
| 10 |
+
{"prompt": "THM ( R v ( R v R ) ) , ( ( S > Q ) > S ) , ( S > Q ) SEQ ( S > ( ( ~ S ) > ( ~ Q ) ) ) PRF", "lean_text": "have n8 : ( R ∨ ( R ∨ R ) ) := h1 ; have n9 : ( ( S → Q ) → S ) := h2 ; have n10 : ( S → ( ( ¬ S ) → ( ¬ Q ) ) ) := ( fun ( n11 : S ) => by have n12 : ( ( ¬ S ) → ( ¬ Q ) ) := ( fun ( n13 : ( ¬ S ) ) => by have n14 : ( ¬ Q ) := ( fun ( n15 : Q ) => by have n16 : False := n13 n11 ; exact ( n16 : False ) ) ; exact n14 ) ; exact n12 ) ; exact n10", "nd": "N1 ( R v ( R v R ) ) : PR ; N2 ( ( S > Q ) > S ) : PR ; N3 | S : AS ; N4 | | ( ~ S ) : AS ; N5 | | | Q : AS ; N6 | | | F : NEGE N3 N4 ; N7 | | ( ~ Q ) : NEGI N5 N6 ; N8 | ( ( ~ S ) > ( ~ Q ) ) : IMPI N4 N7 ; N9 ( S > ( ( ~ S ) > ( ~ Q ) ) ) : IMPI N3 N8 ; QED", "nd_ok": false, "lean_ok": true}
|
| 11 |
+
{"prompt": "THM ( ( S & P ) v S ) , ( ( ( ( S & P ) v S ) v Q ) > Q ) SEQ ( ( ( ( S & P ) v S ) v Q ) > ( S > ( ( ~ ( ~ Q ) ) > Q ) ) ) PRF", "lean_text": "have n45 : ( ( S ∧ P ) ∨ S ) := h1 ; have n46 : ( ( ( ( S ∧ P ) ∨ S ) ∨ Q ) → ( S → ( ( ¬ ( ¬ Q ) ) → Q ) ) ) := ( fun ( n47 : ( ( ( S ∧ P ) ∨ S ) ∨ Q ) ) => by have n48 : ( S → ( ( ¬ ( ¬ Q ) ) → Q ) ) := ( fun ( n49 : S ) => by have n50 : ( ( ¬ ( ¬ Q ) ) → Q ) := ( fun ( n51 : ( ¬ ( ¬ Q ) ) ) => by have n52 : Q := Classical.byContradiction ( fun hh => n51 hh ) ; exact n52 ) ; exact n50 ) ; exact n48 ) ; exact n46", "nd": "N1 ( ( S & P ) v S ) : PR ; N2 | ( ( ( S & P ) v S ) v Q ) : AS ; N3 | | S : AS ; N4 | | | ( ~ ( ~ Q ) ) : AS ; N5 | | | Q : DN N4 ; N6 | | ( ( ~ ( ~ Q ) ) > Q ) : IMPI N4 N5 ; N7 | ( S > ( ( ~ ( ~ Q ) ) > Q ) ) : IMPI N3 N6 ; N8 ( ( ( ( S & P ) v S ) v Q ) > ( S > ( ( ~ ( ~ Q ) ) > Q ) ) ) : IMPI N2 N7 ; QED", "nd_ok": false, "lean_ok": true}
|
| 12 |
+
{"prompt": "THM P , ( ( ~ ( P & P ) ) > Q ) SEQ ( ( ~ ( ~ Q ) ) > ( ( ~ ( ~ S ) ) > ( ( ( ~ ( P & P ) ) > Q ) & P ) ) ) PRF", "lean_text": "have n29 : P := h1 ; have n30 : ( ( ¬ ( ¬ Q ) ) → ( ( ¬ ( ¬ S ) ) → ( ( ( ¬ ( P ∧ P ) ) → Q ) ∧ P ) ) ) := ( fun ( n31 : ( ¬ ( ¬ Q ) ) ) => by have n32 : ( ( ¬ ( ¬ S ) ) → ( ( ( ¬ ( P ∧ P ) ) → Q ) ∧ P ) ) := ( fun ( n33 : ( ¬ ( ¬ S ) ) ) => by have n34 : ( ( ¬ ( P ∧ P ) ) → Q ) := ( fun ( n35 : ( ¬ ( P ∧ P ) ) ) => by have n36 : Q := Classical.byContradiction ( fun hh => n31 hh ) ; exact n36 ) ; have n37 : ( ( ( ¬ ( P ∧ P ) ) → Q ) ∧ P ) := ⟨ n34 , n29 ⟩ ; exact n37 ) ; exact n32 ) ; exact n30", "nd": "N1 P : PR ; N2 | ( ~ ( ~ Q ) ) : AS ; N3 | | ( ~ ( ~ S ) ) : AS ; N4 | | | ( ~ ( P & P ) ) : AS ; N5 | | | Q : DN N2 ; N6 | | ( ( ~ ( P & P ) ) > Q ) : IMPI N4 N5 ; N7 | | ( ( ( ~ ( P & P ) ) > Q ) & P ) : ANDI N6 N1 ; N8 | ( ( ~ ( ~ S ) ) > ( ( ( ~ ( P & P ) ) > Q ) & P ) ) : IMPI N3 N7 ; N9 ( ( ~ ( ~ Q ) ) > ( ( ~ ( ~ S ) ) > ( ( ( ~ ( P & P ) ) > Q ) & P ) ) ) : IMPI N2 N8 ; QED", "nd_ok": false, "lean_ok": true}
|
| 13 |
+
{"prompt": "THM ( ( S & P ) v S ) , ( ( ( ( S & P ) v S ) v Q ) > Q ) SEQ ( ( ( ( S & P ) v S ) v Q ) > ( S > ( ( ~ ( ~ Q ) ) > Q ) ) ) PRF", "lean_text": "have n1 : ( ( S ∧ P ) ∨ S ) := h1 ; have n2 : ( ( ( ( S ∧ P ) ∨ S ) ∨ Q ) → ( S → ( ( ¬ ( ¬ Q ) ) → Q ) ) ) := ( fun ( n3 : ( ( ( S ∧ P ) ∨ S ) ∨ Q ) ) => by have n4 : ( S → ( ( ¬ ( ¬ Q ) ) → Q ) ) := ( fun ( n5 : S ) => by have n6 : ( ( ¬ ( ¬ Q ) ) → Q ) := ( fun ( n7 : ( ¬ ( ¬ Q ) ) ) => by have n8 : Q := Classical.byContradiction ( fun hh => n7 hh ) ; exact n8 ) ; exact n6 ) ; exact n4 ) ; exact n2", "nd": "N1 ( ( S & P ) v S ) : PR ; N2 | ( ( ( S & P ) v S ) v Q ) : AS ; N3 | | S : AS ; N4 | | | ( ~ ( ~ Q ) ) : AS ; N5 | | | Q : DN N4 ; N6 | | ( ( ~ ( ~ Q ) ) > Q ) : IMPI N4 N5 ; N7 | ( S > ( ( ~ ( ~ Q ) ) > Q ) ) : IMPI N3 N6 ; N8 ( ( ( ( S & P ) v S ) v Q ) > ( S > ( ( ~ ( ~ Q ) ) > Q ) ) ) : IMPI N2 N7 ; QED", "nd_ok": false, "lean_ok": true}
|
| 14 |
+
{"prompt": "THM ( ( ~ ( S > S ) ) v ( ~ S ) ) , ( ~ ( ~ ( ~ R ) ) ) , ( ( ~ ( ~ ( ~ R ) ) ) > Q ) SEQ ( ( ~ ( ~ ( ~ R ) ) ) > ( ( P & ( Q v Q ) ) > ( Q v Q ) ) ) PRF", "lean_text": "have n25 : ( ( ¬ ( S → S ) ) ∨ ( ¬ S ) ) := h1 ; have n26 : ( ¬ ( ¬ ( ¬ R ) ) ) := h2 ; have n27 : ( ( ¬ ( ¬ ( ¬ R ) ) ) → ( ( P ∧ ( Q ∨ Q ) ) → ( Q ∨ Q ) ) ) := ( fun ( n28 : ( ¬ ( ¬ ( ¬ R ) ) ) ) => by have n29 : ( ( P ∧ ( Q ∨ Q ) ) → ( Q ∨ Q ) ) := ( fun ( n30 : ( P ∧ ( Q ∨ Q ) ) ) => by have n31 : ( Q ∨ Q ) := n30.2 ; exact n31 ) ; exact n29 ) ; exact n27", "nd": "N1 ( ( ~ ( S > S ) ) v ( ~ S ) ) : PR ; N2 ( ~ ( ~ ( ~ R ) ) ) : PR ; N3 | ( ~ ( ~ ( ~ R ) ) ) : AS ; N4 | | ( P & ( Q v Q ) ) : AS ; N5 | | ( Q v Q ) : ANDE2 N4 ; N6 | ( ( P & ( Q v Q ) ) > ( Q v Q ) ) : IMPI N4 N5 ; N7 ( ( ~ ( ~ ( ~ R ) ) ) > ( ( P & ( Q v Q ) ) > ( Q v Q ) ) ) : IMPI N3 N6 ; QED", "nd_ok": false, "lean_ok": true}
|
| 15 |
+
{"prompt": "THM ( R v ( R v R ) ) , ( ( S > Q ) > S ) , ( S > Q ) SEQ ( S > ( ( ~ S ) > ( ~ Q ) ) ) PRF", "lean_text": "have n50 : ( R ∨ ( R ∨ R ) ) := h1 ; have n51 : ( ( S → Q ) → S ) := h2 ; have n52 : ( S → ( ( ¬ S ) → ( ¬ Q ) ) ) := ( fun ( n53 : S ) => by have n54 : ( ( ¬ S ) → ( ¬ Q ) ) := ( fun ( n55 : ( ¬ S ) ) => by have n56 : ( ¬ Q ) := ( fun ( n57 : Q ) => by have n58 : False := n55 n53 ; exact ( n58 : False ) ) ; exact n56 ) ; exact n54 ) ; exact n52", "nd": "N1 ( R v ( R v R ) ) : PR ; N2 ( ( S > Q ) > S ) : PR ; N3 | S : AS ; N4 | | ( ~ S ) : AS ; N5 | | | Q : AS ; N6 | | | F : NEGE N3 N4 ; N7 | | ( ~ Q ) : NEGI N5 N6 ; N8 | ( ( ~ S ) > ( ~ Q ) ) : IMPI N4 N7 ; N9 ( S > ( ( ~ S ) > ( ~ Q ) ) ) : IMPI N3 N8 ; QED", "nd_ok": false, "lean_ok": true}
|
| 16 |
+
{"prompt": "THM ( ( R & R ) > ( R & S ) ) , ( R & R ) , ( R v P ) SEQ ( ( R & R ) > ( ( R & R ) > R ) ) PRF", "lean_text": "have n40 : ( ( R ∧ R ) → ( R ∧ S ) ) := h1 ; have n41 : ( R ∧ R ) := h2 ; have n42 : ( R ∧ R ) := n41 ; have n43 : ( ( R ∧ R ) → ( ( R ∧ R ) → R ) ) := ( fun ( n44 : ( R ∧ R ) ) => by have n45 : ( ( R ∧ R ) → R ) := ( fun ( n46 : ( R ∧ R ) ) => by have n47 : R := n41.2 ; exact n47 ) ; exact n45 ) ; exact n43", "nd": "N1 ( ( R & R ) > ( R & S ) ) : PR ; N2 ( R & R ) : PR ; N3 ( R & R ) : R N2 ; N4 | ( R & R ) : AS ; N5 | | ( R & R ) : AS ; N6 | | R : ANDE2 N2 ; N7 | ( ( R & R ) > R ) : IMPI N5 N6 ; N8 ( ( R & R ) > ( ( R & R ) > R ) ) : IMPI N4 N7 ; QED", "nd_ok": false, "lean_ok": true}
|
| 17 |
+
{"prompt": "THM ( ~ ( Q > S ) ) , ( S v ( ( P & S ) v ( S v P ) ) ) SEQ ( ~ ( ~ ( ( ~ ( ~ ( Q > S ) ) ) > ( P v ( ( S v ( ( P & S ) v ( S v P ) ) ) & ( Q > S ) ) ) ) ) ) PRF", "lean_text": "have n35 : ( ¬ ( Q → S ) ) := h1 ; have n36 : ( S ∨ ( ( P ∧ S ) ∨ ( S ∨ P ) ) ) := h2 ; have n37 : ( ¬ ( ¬ ( ( ¬ ( ¬ ( Q → S ) ) ) → ( P ∨ ( ( S ∨ ( ( P ∧ S ) ∨ ( S ∨ P ) ) ) ∧ ( Q → S ) ) ) ) ) ) := ( fun ( n38 : ( ¬ ( ( ¬ ( ¬ ( Q → S ) ) ) → ( P ∨ ( ( S ∨ ( ( P ∧ S ) ∨ ( S ∨ P ) ) ) ∧ ( Q → S ) ) ) ) ) ) => by have n39 : ( ( ¬ ( ¬ ( Q → S ) ) ) → ( P ∨ ( ( S ∨ ( ( P ∧ S ) ∨ ( S ∨ P ) ) ) ∧ ( Q → S ) ) ) ) := ( fun ( n40 : ( ¬ ( ¬ ( Q → S ) ) ) ) => by have n41 : ( Q → S ) := Classical.byContradiction ( fun hh => n40 hh ) ; have n42 : ( ( S ∨ ( ( P ∧ S ) ∨ ( S ∨ P ) ) ) ∧ ( Q → S ) ) := ⟨ n36 , n41 ⟩ ; have n43 : ( P ∨ ( ( S ∨ ( ( P ∧ S ) ∨ ( S ∨ P ) ) ) ∧ ( Q → S ) ) ) := Or.inr n42 ; exact n43 ) ; have n44 : False := n38 n39 ; exact n44 ) ; exact n37", "nd": "N1 ( ~ ( Q > S ) ) : PR ; N2 ( S v ( ( P & S ) v ( S v P ) ) ) : PR ; N3 | ( ~ ( ( ~ ( ~ ( Q > S ) ) ) > ( P v ( ( S v ( ( P & S ) v ( S v P ) ) ) & ( Q > S ) ) ) ) ) : AS ; N4 | | ( ~ ( ~ ( Q > S ) ) ) : AS ; N5 | | ( Q > S ) : DN N4 ; N6 | | ( ( S v ( ( P & S ) v ( S v P ) ) ) & ( Q > S ) ) : ANDI N2 N5 ; N7 | | ( P v ( ( S v ( ( P & S ) v ( S v P ) ) ) & ( Q > S ) ) ) : ORI2 N6 ; N8 | ( ( ~ ( ~ ( Q > S ) ) ) > ( P v ( ( S v ( ( P & S ) v ( S v P ) ) ) & ( Q > S ) ) ) ) : IMPI N4 N7 ; N9 | F : NEGE N8 N3 ; N10 ( ~ ( ~ ( ( ~ ( ~ ( Q > S ) ) ) > ( P v ( ( S v ( ( P & S ) v ( S v P ) ) ) & ( Q > S ) ) ) ) ) ) : IMPI N3 N9 ; QED", "nd_ok": false, "lean_ok": true}
|
| 18 |
+
{"prompt": "THM ( ( ~ ( ~ S ) ) v Q ) SEQ ( ( ~ ( ~ S ) ) > ( P > ( ( ~ S ) > ( ( ( ~ ( ~ S ) ) v Q ) & ( ( ~ ( ~ S ) ) v Q ) ) ) ) ) PRF", "lean_text": "have n44 : ( ( ¬ ( ¬ S ) ) → ( P → ( ( ¬ S ) → ( ( ( ¬ ( ¬ S ) ) ∨ Q ) ∧ ( ( ¬ ( ¬ S ) ) ∨ Q ) ) ) ) ) := ( fun ( n45 : ( ¬ ( ¬ S ) ) ) => by have n46 : ( ( ¬ ( ¬ S ) ) ∨ Q ) := Or.inl n45 ; have n47 : ( P → ( ( ¬ S ) → ( ( ( ¬ ( ¬ S ) ) ∨ Q ) ∧ ( ( ¬ ( ¬ S ) ) ∨ Q ) ) ) ) := ( fun ( n48 : P ) => by have n49 : ( ( ¬ ( ¬ S ) ) ∨ Q ) := Or.inl n45 ; have n50 : ( ( ( ¬ ( ¬ S ) ) ∨ Q ) ∧ ( ( ¬ ( ¬ S ) ) ∨ Q ) ) := ⟨ n49 , n49 ⟩ ; have n51 : ( ( ¬ S ) → ( ( ( ¬ ( ¬ S ) ) ∨ Q ) ∧ ( ( ¬ ( ¬ S ) ) ∨ Q ) ) ) := ( fun ( n52 : ( ¬ S ) ) => by have n53 : ( ( ( ¬ ( ¬ S ) ) ∨ Q ) ∧ ( ( ¬ ( ¬ S ) ) ∨ Q ) ) := ⟨ n49 , n49 ⟩ ; exact n53 ) ; exact n51 ) ; exact n47 ) ; exact n44", "nd": "N1 | ( ~ ( ~ S ) ) : AS ; N2 | ( ( ~ ( ~ S ) ) v Q ) : ORI1 N1 ; N3 | | P : AS ; N4 | | ( ( ~ ( ~ S ) ) v Q ) : ORI1 N1 ; N5 | | ( ( ( ~ ( ~ S ) ) v Q ) & ( ( ~ ( ~ S ) ) v Q ) ) : ANDI N4 N4 ; N6 | | | ( ~ S ) : AS ; N7 | | | ( ( ( ~ ( ~ S ) ) v Q ) & ( ( ~ ( ~ S ) ) v Q ) ) : ANDI N4 N4 ; N8 | | ( ( ~ S ) > ( ( ( ~ ( ~ S ) ) v Q ) & ( ( ~ ( ~ S ) ) v Q ) ) ) : IMPI N6 N7 ; N9 | ( P > ( ( ~ S ) > ( ( ( ~ ( ~ S ) ) v Q ) & ( ( ~ ( ~ S ) ) v Q ) ) ) ) : IMPI N3 N8 ; N10 ( ( ~ ( ~ S ) ) > ( P > ( ( ~ S ) > ( ( ( ~ ( ~ S ) ) v Q ) & ( ( ~ ( ~ S ) ) v Q ) ) ) ) ) : IMPI N1 N9 ; QED", "nd_ok": false, "lean_ok": true}
|
| 19 |
+
{"prompt": "THM ( ( P > S ) & P ) , P SEQ ( ( ~ ( ~ ( ~ R ) ) ) > ( ( ~ ( ~ S ) ) > ( R > S ) ) ) PRF", "lean_text": "have n37 : ( ( P → S ) ∧ P ) := h1 ; have n38 : P := n37.2 ; have n39 : ( ( ¬ ( ¬ ( ¬ R ) ) ) → ( ( ¬ ( ¬ S ) ) → ( R → S ) ) ) := ( fun ( n40 : ( ¬ ( ¬ ( ¬ R ) ) ) ) => by have n41 : ( ( ¬ ( ¬ S ) ) → ( R → S ) ) := ( fun ( n42 : ( ¬ ( ¬ S ) ) ) => by have n43 : ( P → S ) := n37.1 ; have n44 : ( R → S ) := ( fun ( n45 : R ) => by have n46 : S := n43 n38 ; exact n46 ) ; exact n44 ) ; exact n41 ) ; exact n39", "nd": "N1 ( ( P > S ) & P ) : PR ; N2 P : ANDE2 N1 ; N3 | ( ~ ( ~ ( ~ R ) ) ) : AS ; N4 | | ( ~ ( ~ S ) ) : AS ; N5 | | ( P > S ) : ANDE1 N1 ; N6 | | | R : AS ; N7 | | | S : IMPE N5 N2 ; N8 | | ( R > S ) : IMPI N6 N7 ; N9 | ( ( ~ ( ~ S ) ) > ( R > S ) ) : IMPI N4 N8 ; N10 ( ( ~ ( ~ ( ~ R ) ) ) > ( ( ~ ( ~ S ) ) > ( R > S ) ) ) : IMPI N3 N9 ; QED", "nd_ok": false, "lean_ok": true}
|
runs/510-fable_more-shorter-rounds_s1/lean_gate.jsonl
ADDED
|
@@ -0,0 +1,60 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{"utc": "2026-10-07T23:31:26Z", "samples": 2000, "parse_fail": 6, "distinct_checked": 1994, "both_ok": 1920, "nd_ok_lean_rej": 0, "nd_rej_lean_ok": 1, "both_rej": 73, "parse_reasons": {"unbound": 5, "formula": 1}, "lean_wall_s": 3.594557285308838, "lean_proc_s": 16.63570761680603, "nd_verify_s": 0.10666513442993164, "workers": 96, "chunk": 400}
|
| 2 |
+
{"utc": "2026-10-07T23:32:17Z", "samples": 12000, "parse_fail": 5002, "distinct_checked": 6801, "both_ok": 1285, "nd_ok_lean_rej": 0, "nd_rej_lean_ok": 4, "both_rej": 5512, "parse_reasons": {"unbound": 3444, "formula": 656, "expected )": 232, "exact not last": 154, "expected (": 143, "name": 119, "trailing tokens": 81, "no-eos": 59, "expected :=": 41, "PR position": 38, "expected ;": 29, "ORE box": 5, "expected =>": 1}, "lean_wall_s": 23.476678609848022, "lean_proc_s": 396.88426542282104, "nd_verify_s": 0.37419891357421875, "workers": 96, "chunk": 400}
|
| 3 |
+
{"utc": "2026-10-07T23:33:02Z", "samples": 8760, "parse_fail": 3133, "distinct_checked": 5165, "both_ok": 515, "nd_ok_lean_rej": 0, "nd_rej_lean_ok": 0, "both_rej": 4650, "parse_reasons": {"unbound": 2167, "formula": 353, "expected (": 135, "name": 133, "expected )": 109, "exact not last": 74, "expected :=": 64, "trailing tokens": 34, "PR position": 30, "ORE box": 16, "no-eos": 11, "expected None": 3, "expected ;": 3, "expected .elim": 1}, "lean_wall_s": 20.264501333236694, "lean_proc_s": 261.473201751709, "nd_verify_s": 0.27551960945129395, "workers": 96, "chunk": 400}
|
| 4 |
+
{"utc": "2026-10-07T23:33:59Z", "samples": 8760, "parse_fail": 2879, "distinct_checked": 5369, "both_ok": 855, "nd_ok_lean_rej": 0, "nd_rej_lean_ok": 0, "both_rej": 4514, "parse_reasons": {"unbound": 1934, "formula": 204, "expected (": 185, "name": 174, "expected )": 95, "exact not last": 84, "no-eos": 59, "expected :=": 57, "PR position": 55, "expected ;": 11, "trailing tokens": 10, "ORE box": 7, "expected =>": 3, "expected :": 1}, "lean_wall_s": 21.655895709991455, "lean_proc_s": 291.3210391998291, "nd_verify_s": 0.29367613792419434, "workers": 96, "chunk": 400}
|
| 5 |
+
{"utc": "2026-10-07T23:35:22Z", "samples": 8736, "parse_fail": 2723, "distinct_checked": 5486, "both_ok": 1010, "nd_ok_lean_rej": 0, "nd_rej_lean_ok": 0, "both_rej": 4476, "parse_reasons": {"unbound": 1890, "formula": 243, "name": 173, "expected (": 122, "expected )": 106, "exact not last": 67, "PR position": 36, "expected :=": 28, "ORE box": 20, "trailing tokens": 15, "expected ;": 8, "no-eos": 7, "expected =>": 2, "expected by": 2, "expected None": 2, "expected .elim": 1, "expected exact": 1}, "lean_wall_s": 21.18886423110962, "lean_proc_s": 289.83601808547974, "nd_verify_s": 0.300525426864624, "workers": 96, "chunk": 400}
|
| 6 |
+
{"utc": "2026-10-07T23:36:45Z", "samples": 12000, "parse_fail": 1614, "distinct_checked": 10020, "both_ok": 4687, "nd_ok_lean_rej": 0, "nd_rej_lean_ok": 0, "both_rej": 5333, "parse_reasons": {"unbound": 1196, "formula": 126, "name": 66, "expected (": 62, "no-eos": 51, "expected )": 46, "exact not last": 26, "expected :=": 14, "ORE box": 10, "PR position": 9, "expected ;": 4, "expected =>": 4}, "lean_wall_s": 30.24298596382141, "lean_proc_s": 746.648366689682, "nd_verify_s": 0.5275154113769531, "workers": 96, "chunk": 400}
|
| 7 |
+
{"utc": "2026-10-07T23:37:16Z", "samples": 5448, "parse_fail": 1253, "distinct_checked": 3843, "both_ok": 225, "nd_ok_lean_rej": 0, "nd_rej_lean_ok": 0, "both_rej": 3618, "parse_reasons": {"unbound": 910, "expected (": 107, "formula": 106, "name": 55, "no-eos": 27, "PR position": 13, "expected )": 13, "exact not last": 7, "expected :=": 6, "expected None": 4, "ORE box": 2, "expected ;": 2, "trailing tokens": 1}, "lean_wall_s": 16.1001238822937, "lean_proc_s": 155.76910042762756, "nd_verify_s": 0.22156691551208496, "workers": 96, "chunk": 400}
|
| 8 |
+
{"utc": "2026-10-07T23:37:54Z", "samples": 5448, "parse_fail": 1209, "distinct_checked": 3894, "both_ok": 464, "nd_ok_lean_rej": 0, "nd_rej_lean_ok": 0, "both_rej": 3430, "parse_reasons": {"unbound": 976, "expected (": 73, "name": 55, "formula": 24, "PR position": 20, "exact not last": 13, "ORE box": 11, "expected )": 10, "expected :=": 9, "expected ;": 9, "no-eos": 5, "trailing tokens": 2, "expected None": 1, "expected =>": 1}, "lean_wall_s": 16.533779859542847, "lean_proc_s": 160.46765208244324, "nd_verify_s": 0.2295055389404297, "workers": 96, "chunk": 400}
|
| 9 |
+
{"utc": "2026-10-07T23:38:43Z", "samples": 5424, "parse_fail": 1022, "distinct_checked": 4025, "both_ok": 585, "nd_ok_lean_rej": 0, "nd_rej_lean_ok": 0, "both_rej": 3440, "parse_reasons": {"unbound": 874, "name": 56, "expected (": 25, "formula": 25, "PR position": 14, "exact not last": 9, "expected ;": 5, "expected False": 4, "expected )": 3, "no-eos": 3, "ORE box": 2, "trailing tokens": 1, "expected None": 1}, "lean_wall_s": 15.78660273551941, "lean_proc_s": 157.20841145515442, "nd_verify_s": 0.2511265277862549, "workers": 96, "chunk": 400}
|
| 10 |
+
{"utc": "2026-10-07T23:40:07Z", "samples": 12000, "parse_fail": 1570, "distinct_checked": 10103, "both_ok": 5710, "nd_ok_lean_rej": 0, "nd_rej_lean_ok": 0, "both_rej": 4393, "parse_reasons": {"unbound": 1066, "formula": 152, "expected )": 106, "expected (": 70, "name": 55, "no-eos": 45, "exact not last": 23, "expected ;": 16, "expected :=": 15, "ORE box": 8, "trailing tokens": 7, "expected None": 3, "expected =>": 3, "PR position": 1}, "lean_wall_s": 27.114699125289917, "lean_proc_s": 676.2209341526031, "nd_verify_s": 0.541231632232666, "workers": 96, "chunk": 400}
|
| 11 |
+
{"utc": "2026-10-07T23:40:32Z", "samples": 4344, "parse_fail": 1004, "distinct_checked": 3095, "both_ok": 284, "nd_ok_lean_rej": 0, "nd_rej_lean_ok": 0, "both_rej": 2811, "parse_reasons": {"unbound": 751, "expected )": 65, "formula": 42, "expected (": 38, "name": 31, "ORE box": 30, "expected ;": 22, "expected :=": 8, "exact not last": 8, "no-eos": 5, "trailing tokens": 2, "expected =>": 1, "expected None": 1}, "lean_wall_s": 11.517638683319092, "lean_proc_s": 90.15449118614197, "nd_verify_s": 0.17023491859436035, "workers": 96, "chunk": 400}
|
| 12 |
+
{"utc": "2026-10-07T23:41:01Z", "samples": 4344, "parse_fail": 1309, "distinct_checked": 2812, "both_ok": 414, "nd_ok_lean_rej": 0, "nd_rej_lean_ok": 0, "both_rej": 2398, "parse_reasons": {"unbound": 1012, "name": 88, "expected (": 56, "formula": 42, "expected )": 38, "expected ;": 33, "expected :=": 21, "ORE box": 11, "PR position": 3, "no-eos": 2, "exact not last": 2, "expected False": 1}, "lean_wall_s": 8.148078680038452, "lean_proc_s": 55.42657661437988, "nd_verify_s": 0.14136958122253418, "workers": 96, "chunk": 400}
|
| 13 |
+
{"utc": "2026-10-07T23:41:41Z", "samples": 4320, "parse_fail": 1025, "distinct_checked": 3019, "both_ok": 475, "nd_ok_lean_rej": 0, "nd_rej_lean_ok": 1, "both_rej": 2543, "parse_reasons": {"unbound": 737, "expected (": 85, "formula": 43, "name": 42, "expected )": 36, "expected :=": 28, "no-eos": 19, "exact not last": 12, "expected ;": 10, "PR position": 8, "expected =>": 2, "expected None": 1, "expected False": 1, "expected ,": 1}, "lean_wall_s": 9.161462306976318, "lean_proc_s": 68.29506397247314, "nd_verify_s": 0.17038440704345703, "workers": 96, "chunk": 400}
|
| 14 |
+
{"utc": "2026-10-07T23:43:09Z", "samples": 12000, "parse_fail": 1631, "distinct_checked": 10064, "both_ok": 6251, "nd_ok_lean_rej": 0, "nd_rej_lean_ok": 0, "both_rej": 3813, "parse_reasons": {"unbound": 1067, "formula": 203, "name": 85, "no-eos": 71, "expected (": 52, "expected ;": 49, "expected )": 45, "exact not last": 31, "expected :=": 15, "ORE box": 6, "trailing tokens": 2, "expected False": 2, "expected None": 2, "PR position": 1}, "lean_wall_s": 28.171414613723755, "lean_proc_s": 700.5533065795898, "nd_verify_s": 0.6327154636383057, "workers": 96, "chunk": 400}
|
| 15 |
+
{"utc": "2026-10-07T23:43:27Z", "samples": 3528, "parse_fail": 994, "distinct_checked": 2357, "both_ok": 221, "nd_ok_lean_rej": 0, "nd_rej_lean_ok": 1, "both_rej": 2135, "parse_reasons": {"unbound": 729, "name": 82, "formula": 47, "expected )": 45, "expected :=": 30, "expected ;": 21, "expected (": 21, "exact not last": 9, "ORE box": 4, "trailing tokens": 2, "expected exact": 1, "PR position": 1, "no-eos": 1, "expected None": 1}, "lean_wall_s": 6.499085903167725, "lean_proc_s": 37.98618745803833, "nd_verify_s": 0.12749075889587402, "workers": 96, "chunk": 400}
|
| 16 |
+
{"utc": "2026-10-07T23:43:50Z", "samples": 3528, "parse_fail": 957, "distinct_checked": 2369, "both_ok": 263, "nd_ok_lean_rej": 0, "nd_rej_lean_ok": 0, "both_rej": 2106, "parse_reasons": {"unbound": 710, "name": 53, "formula": 50, "expected (": 38, "no-eos": 25, "expected :=": 22, "expected )": 21, "expected ;": 20, "exact not last": 10, "PR position": 7, "trailing tokens": 1}, "lean_wall_s": 7.598954200744629, "lean_proc_s": 44.85918092727661, "nd_verify_s": 0.1153712272644043, "workers": 96, "chunk": 400}
|
| 17 |
+
{"utc": "2026-10-07T23:44:35Z", "samples": 3504, "parse_fail": 914, "distinct_checked": 2408, "both_ok": 497, "nd_ok_lean_rej": 0, "nd_rej_lean_ok": 0, "both_rej": 1911, "parse_reasons": {"unbound": 697, "expected (": 56, "name": 47, "no-eos": 35, "formula": 25, "expected )": 19, "exact not last": 13, "expected ;": 10, "expected :=": 7, "PR position": 3, "expected =>": 1, "ORE box": 1}, "lean_wall_s": 7.9225099086761475, "lean_proc_s": 47.64143753051758, "nd_verify_s": 0.12872719764709473, "workers": 96, "chunk": 400}
|
| 18 |
+
{"utc": "2026-10-07T23:46:06Z", "samples": 12000, "parse_fail": 1365, "distinct_checked": 10259, "both_ok": 6642, "nd_ok_lean_rej": 0, "nd_rej_lean_ok": 4, "both_rej": 3613, "parse_reasons": {"unbound": 936, "formula": 130, "name": 89, "expected (": 88, "expected )": 39, "no-eos": 26, "exact not last": 22, "expected ;": 9, "ORE box": 8, "expected :=": 8, "trailing tokens": 6, "PR position": 4}, "lean_wall_s": 31.011502027511597, "lean_proc_s": 789.0329129695892, "nd_verify_s": 0.5945971012115479, "workers": 96, "chunk": 400}
|
| 19 |
+
{"utc": "2026-10-07T23:46:26Z", "samples": 3144, "parse_fail": 1030, "distinct_checked": 1975, "both_ok": 201, "nd_ok_lean_rej": 0, "nd_rej_lean_ok": 0, "both_rej": 1774, "parse_reasons": {"unbound": 644, "expected (": 114, "formula": 79, "no-eos": 52, "name": 48, "expected :=": 34, "expected )": 26, "ORE box": 12, "trailing tokens": 9, "exact not last": 6, "PR position": 3, "expected ;": 2, "expected None": 1}, "lean_wall_s": 5.465736627578735, "lean_proc_s": 26.949828147888184, "nd_verify_s": 0.12272763252258301, "workers": 96, "chunk": 400}
|
| 20 |
+
{"utc": "2026-10-07T23:46:52Z", "samples": 3120, "parse_fail": 1027, "distinct_checked": 1952, "both_ok": 361, "nd_ok_lean_rej": 0, "nd_rej_lean_ok": 0, "both_rej": 1591, "parse_reasons": {"unbound": 729, "expected (": 116, "name": 78, "formula": 32, "expected )": 28, "expected :=": 24, "expected ;": 8, "exact not last": 4, "ORE box": 4, "no-eos": 2, "PR position": 1, "expected =>": 1}, "lean_wall_s": 7.6350579261779785, "lean_proc_s": 37.05641508102417, "nd_verify_s": 0.09786534309387207, "workers": 96, "chunk": 400}
|
| 21 |
+
{"utc": "2026-10-07T23:47:30Z", "samples": 3120, "parse_fail": 906, "distinct_checked": 2031, "both_ok": 348, "nd_ok_lean_rej": 0, "nd_rej_lean_ok": 0, "both_rej": 1683, "parse_reasons": {"unbound": 612, "expected (": 86, "name": 71, "formula": 50, "expected :=": 25, "expected )": 21, "no-eos": 21, "ORE box": 12, "exact not last": 4, "expected ;": 3, "expected =>": 1}, "lean_wall_s": 7.414982080459595, "lean_proc_s": 37.6883430480957, "nd_verify_s": 0.12455177307128906, "workers": 96, "chunk": 400}
|
| 22 |
+
{"utc": "2026-10-07T23:48:53Z", "samples": 12000, "parse_fail": 1395, "distinct_checked": 10272, "both_ok": 6921, "nd_ok_lean_rej": 0, "nd_rej_lean_ok": 2, "both_rej": 3349, "parse_reasons": {"unbound": 1111, "formula": 100, "expected (": 59, "name": 44, "exact not last": 23, "expected )": 22, "expected ;": 21, "no-eos": 5, "ORE box": 4, "trailing tokens": 2, "expected =>": 1, "PR position": 1, "expected :=": 1, "expected None": 1}, "lean_wall_s": 24.084243297576904, "lean_proc_s": 611.9154925346375, "nd_verify_s": 0.623915433883667, "workers": 96, "chunk": 400}
|
| 23 |
+
{"utc": "2026-10-07T23:49:10Z", "samples": 2640, "parse_fail": 1006, "distinct_checked": 1537, "both_ok": 169, "nd_ok_lean_rej": 0, "nd_rej_lean_ok": 0, "both_rej": 1368, "parse_reasons": {"unbound": 851, "formula": 53, "name": 33, "expected (": 26, "expected ;": 14, "exact not last": 11, "expected )": 8, "no-eos": 7, "PR position": 1, "expected :=": 1, "trailing tokens": 1}, "lean_wall_s": 4.416697978973389, "lean_proc_s": 17.180217504501343, "nd_verify_s": 0.07572078704833984, "workers": 96, "chunk": 400}
|
| 24 |
+
{"utc": "2026-10-07T23:49:25Z", "samples": 2640, "parse_fail": 881, "distinct_checked": 1637, "both_ok": 284, "nd_ok_lean_rej": 0, "nd_rej_lean_ok": 0, "both_rej": 1353, "parse_reasons": {"unbound": 707, "expected (": 65, "formula": 36, "name": 28, "expected ;": 17, "expected )": 16, "ORE box": 6, "exact not last": 4, "PR position": 1, "no-eos": 1}, "lean_wall_s": 4.604932069778442, "lean_proc_s": 18.742263793945312, "nd_verify_s": 0.10993671417236328, "workers": 96, "chunk": 400}
|
| 25 |
+
{"utc": "2026-10-07T23:49:54Z", "samples": 2616, "parse_fail": 848, "distinct_checked": 1655, "both_ok": 287, "nd_ok_lean_rej": 0, "nd_rej_lean_ok": 0, "both_rej": 1368, "parse_reasons": {"unbound": 692, "expected (": 47, "name": 29, "expected :=": 23, "formula": 20, "exact not last": 10, "expected )": 10, "ORE box": 9, "expected ;": 6, "expected =>": 1, "no-eos": 1}, "lean_wall_s": 6.155207395553589, "lean_proc_s": 24.88498616218567, "nd_verify_s": 0.08916354179382324, "workers": 96, "chunk": 400}
|
| 26 |
+
{"utc": "2026-10-07T23:51:26Z", "samples": 12000, "parse_fail": 1757, "distinct_checked": 9923, "both_ok": 7269, "nd_ok_lean_rej": 0, "nd_rej_lean_ok": 3, "both_rej": 2651, "parse_reasons": {"unbound": 1185, "formula": 183, "expected (": 182, "name": 70, "no-eos": 61, "exact not last": 26, "expected )": 25, "ORE box": 11, "trailing tokens": 5, "expected :=": 3, "PR position": 3, "expected None": 1, "expected ;": 1, "expected =>": 1}, "lean_wall_s": 29.50361657142639, "lean_proc_s": 722.0783979892731, "nd_verify_s": 0.6333663463592529, "workers": 96, "chunk": 400}
|
| 27 |
+
{"utc": "2026-10-07T23:51:43Z", "samples": 2448, "parse_fail": 995, "distinct_checked": 1347, "both_ok": 230, "nd_ok_lean_rej": 0, "nd_rej_lean_ok": 0, "both_rej": 1117, "parse_reasons": {"unbound": 713, "name": 70, "expected (": 66, "no-eos": 49, "formula": 46, "expected )": 17, "expected ;": 9, "expected =>": 7, "ORE box": 5, "exact not last": 5, "PR position": 5, "expected None": 2, "trailing tokens": 1}, "lean_wall_s": 5.9002721309661865, "lean_proc_s": 20.916486978530884, "nd_verify_s": 0.0765068531036377, "workers": 96, "chunk": 400}
|
| 28 |
+
{"utc": "2026-10-07T23:52:05Z", "samples": 2448, "parse_fail": 890, "distinct_checked": 1439, "both_ok": 262, "nd_ok_lean_rej": 0, "nd_rej_lean_ok": 0, "both_rej": 1177, "parse_reasons": {"unbound": 640, "expected (": 91, "name": 57, "no-eos": 30, "formula": 22, "expected )": 19, "ORE box": 16, "expected ;": 7, "exact not last": 4, "PR position": 2, "expected =>": 1, "expected None": 1}, "lean_wall_s": 5.5459723472595215, "lean_proc_s": 20.586273670196533, "nd_verify_s": 0.07504940032958984, "workers": 96, "chunk": 400}
|
| 29 |
+
{"utc": "2026-10-07T23:52:38Z", "samples": 2424, "parse_fail": 1016, "distinct_checked": 1324, "both_ok": 286, "nd_ok_lean_rej": 0, "nd_rej_lean_ok": 0, "both_rej": 1038, "parse_reasons": {"unbound": 784, "expected (": 78, "name": 46, "ORE box": 30, "no-eos": 26, "formula": 23, "expected )": 17, "PR position": 6, "exact not last": 4, "expected =>": 1, "expected ;": 1}, "lean_wall_s": 5.226251125335693, "lean_proc_s": 17.53129744529724, "nd_verify_s": 0.07537126541137695, "workers": 96, "chunk": 400}
|
| 30 |
+
{"utc": "2026-10-07T23:54:15Z", "samples": 12000, "parse_fail": 1623, "distinct_checked": 10127, "both_ok": 7113, "nd_ok_lean_rej": 0, "nd_rej_lean_ok": 0, "both_rej": 3014, "parse_reasons": {"unbound": 1162, "formula": 115, "name": 67, "no-eos": 61, "exact not last": 51, "expected (": 51, "expected :=": 34, "expected )": 31, "expected ;": 24, "trailing tokens": 13, "ORE box": 7, "PR position": 5, "expected =>": 1, "expected False": 1}, "lean_wall_s": 28.499887466430664, "lean_proc_s": 713.0135202407837, "nd_verify_s": 0.6211800575256348, "workers": 96, "chunk": 400}
|
| 31 |
+
{"utc": "2026-10-07T23:54:31Z", "samples": 2280, "parse_fail": 957, "distinct_checked": 1240, "both_ok": 263, "nd_ok_lean_rej": 0, "nd_rej_lean_ok": 0, "both_rej": 977, "parse_reasons": {"unbound": 721, "name": 84, "formula": 48, "expected )": 27, "expected (": 24, "no-eos": 19, "expected :=": 15, "exact not last": 8, "PR position": 6, "expected ;": 5}, "lean_wall_s": 4.8404035568237305, "lean_proc_s": 15.175577878952026, "nd_verify_s": 0.11170649528503418, "workers": 96, "chunk": 400}
|
| 32 |
+
{"utc": "2026-10-07T23:54:51Z", "samples": 2280, "parse_fail": 740, "distinct_checked": 1444, "both_ok": 316, "nd_ok_lean_rej": 0, "nd_rej_lean_ok": 0, "both_rej": 1128, "parse_reasons": {"unbound": 587, "name": 47, "expected ;": 22, "expected (": 19, "expected )": 19, "expected :=": 16, "formula": 10, "no-eos": 9, "PR position": 5, "exact not last": 3, "ORE box": 3}, "lean_wall_s": 5.603620767593384, "lean_proc_s": 21.297504663467407, "nd_verify_s": 0.07175111770629883, "workers": 96, "chunk": 400}
|
| 33 |
+
{"utc": "2026-10-07T23:55:18Z", "samples": 2280, "parse_fail": 892, "distinct_checked": 1286, "both_ok": 251, "nd_ok_lean_rej": 0, "nd_rej_lean_ok": 0, "both_rej": 1035, "parse_reasons": {"unbound": 666, "name": 60, "expected (": 49, "formula": 39, "ORE box": 24, "expected )": 18, "no-eos": 16, "expected :=": 16, "expected ;": 2, "PR position": 1, "exact not last": 1}, "lean_wall_s": 4.995470285415649, "lean_proc_s": 16.531874656677246, "nd_verify_s": 0.07354235649108887, "workers": 96, "chunk": 400}
|
| 34 |
+
{"utc": "2026-10-07T23:56:52Z", "samples": 12000, "parse_fail": 1559, "distinct_checked": 10230, "both_ok": 7607, "nd_ok_lean_rej": 0, "nd_rej_lean_ok": 0, "both_rej": 2623, "parse_reasons": {"unbound": 1045, "formula": 153, "name": 96, "expected (": 94, "no-eos": 49, "expected )": 40, "exact not last": 33, "ORE box": 21, "expected ;": 15, "expected :=": 9, "expected =>": 3, "expected :": 1}, "lean_wall_s": 27.71985960006714, "lean_proc_s": 704.4079508781433, "nd_verify_s": 0.628711462020874, "workers": 96, "chunk": 400}
|
| 35 |
+
{"utc": "2026-10-07T23:57:03Z", "samples": 2064, "parse_fail": 866, "distinct_checked": 1120, "both_ok": 151, "nd_ok_lean_rej": 0, "nd_rej_lean_ok": 0, "both_rej": 969, "parse_reasons": {"unbound": 555, "name": 85, "expected (": 84, "formula": 47, "ORE box": 33, "expected )": 31, "no-eos": 17, "expected ;": 10, "exact not last": 3, "expected =>": 1}, "lean_wall_s": 2.001866102218628, "lean_proc_s": 5.723641395568848, "nd_verify_s": 0.04319953918457031, "workers": 96, "chunk": 400}
|
| 36 |
+
{"utc": "2026-10-07T23:57:23Z", "samples": 2064, "parse_fail": 893, "distinct_checked": 1093, "both_ok": 146, "nd_ok_lean_rej": 0, "nd_rej_lean_ok": 0, "both_rej": 947, "parse_reasons": {"unbound": 589, "expected (": 133, "name": 83, "expected )": 33, "ORE box": 26, "formula": 14, "no-eos": 5, "exact not last": 5, "PR position": 4, "expected :=": 1}, "lean_wall_s": 4.432354211807251, "lean_proc_s": 12.72371792793274, "nd_verify_s": 0.08528757095336914, "workers": 96, "chunk": 400}
|
| 37 |
+
{"utc": "2026-10-07T23:57:54Z", "samples": 2040, "parse_fail": 796, "distinct_checked": 1165, "both_ok": 197, "nd_ok_lean_rej": 0, "nd_rej_lean_ok": 0, "both_rej": 968, "parse_reasons": {"unbound": 591, "name": 53, "expected (": 47, "formula": 32, "expected )": 27, "no-eos": 22, "expected :=": 13, "exact not last": 6, "expected ;": 2, "trailing tokens": 1, "PR position": 1, "ORE box": 1}, "lean_wall_s": 5.335719347000122, "lean_proc_s": 15.741442441940308, "nd_verify_s": 0.07942891120910645, "workers": 96, "chunk": 400}
|
| 38 |
+
{"utc": "2026-10-07T23:59:38Z", "samples": 12000, "parse_fail": 1422, "distinct_checked": 10309, "both_ok": 7737, "nd_ok_lean_rej": 0, "nd_rej_lean_ok": 0, "both_rej": 2572, "parse_reasons": {"unbound": 908, "name": 102, "no-eos": 89, "expected (": 88, "formula": 88, "exact not last": 43, "expected ;": 33, "expected )": 24, "expected :=": 22, "ORE box": 11, "trailing tokens": 10, "PR position": 2, "expected None": 1, "expected False": 1}, "lean_wall_s": 30.53583574295044, "lean_proc_s": 772.3464777469635, "nd_verify_s": 0.7300162315368652, "workers": 96, "chunk": 400}
|
| 39 |
+
{"utc": "2026-10-07T23:59:52Z", "samples": 2016, "parse_fail": 799, "distinct_checked": 1149, "both_ok": 168, "nd_ok_lean_rej": 0, "nd_rej_lean_ok": 0, "both_rej": 981, "parse_reasons": {"unbound": 501, "expected (": 71, "no-eos": 68, "name": 61, "formula": 34, "expected )": 20, "expected :=": 18, "trailing tokens": 8, "exact not last": 8, "expected ;": 7, "ORE box": 2, "PR position": 1}, "lean_wall_s": 4.629773855209351, "lean_proc_s": 13.672598838806152, "nd_verify_s": 0.06526947021484375, "workers": 96, "chunk": 400}
|
| 40 |
+
{"utc": "2026-10-08T00:00:11Z", "samples": 2016, "parse_fail": 918, "distinct_checked": 1026, "both_ok": 163, "nd_ok_lean_rej": 0, "nd_rej_lean_ok": 0, "both_rej": 863, "parse_reasons": {"unbound": 538, "expected (": 143, "name": 96, "no-eos": 62, "ORE box": 27, "expected ;": 17, "formula": 16, "expected )": 13, "exact not last": 5, "expected None": 1}, "lean_wall_s": 4.028652667999268, "lean_proc_s": 10.97902536392212, "nd_verify_s": 0.04991626739501953, "workers": 96, "chunk": 400}
|
| 41 |
+
{"utc": "2026-10-08T00:00:38Z", "samples": 1992, "parse_fail": 744, "distinct_checked": 1179, "both_ok": 183, "nd_ok_lean_rej": 0, "nd_rej_lean_ok": 0, "both_rej": 996, "parse_reasons": {"unbound": 454, "name": 116, "expected (": 62, "expected ;": 60, "no-eos": 24, "ORE box": 10, "expected )": 7, "formula": 6, "exact not last": 3, "PR position": 2}, "lean_wall_s": 4.079252004623413, "lean_proc_s": 11.83238697052002, "nd_verify_s": 0.05672812461853027, "workers": 96, "chunk": 400}
|
| 42 |
+
{"utc": "2026-10-08T00:02:15Z", "samples": 12000, "parse_fail": 1458, "distinct_checked": 10323, "both_ok": 7813, "nd_ok_lean_rej": 0, "nd_rej_lean_ok": 0, "both_rej": 2510, "parse_reasons": {"unbound": 1041, "formula": 104, "name": 82, "expected (": 65, "no-eos": 56, "exact not last": 42, "trailing tokens": 24, "expected )": 16, "ORE box": 12, "expected ;": 11, "expected :=": 3, "expected hh": 1, "expected False": 1}, "lean_wall_s": 25.536564588546753, "lean_proc_s": 651.7314283847809, "nd_verify_s": 0.666837215423584, "workers": 96, "chunk": 400}
|
| 43 |
+
{"utc": "2026-10-08T00:02:28Z", "samples": 1872, "parse_fail": 717, "distinct_checked": 1086, "both_ok": 77, "nd_ok_lean_rej": 0, "nd_rej_lean_ok": 0, "both_rej": 1009, "parse_reasons": {"unbound": 504, "name": 76, "expected (": 50, "formula": 30, "no-eos": 26, "expected )": 9, "exact not last": 8, "ORE box": 5, "expected ;": 4, "expected :=": 3, "PR position": 1, "expected None": 1}, "lean_wall_s": 3.4952971935272217, "lean_proc_s": 10.039034128189087, "nd_verify_s": 0.0519719123840332, "workers": 96, "chunk": 400}
|
| 44 |
+
{"utc": "2026-10-08T00:02:43Z", "samples": 1872, "parse_fail": 895, "distinct_checked": 935, "both_ok": 103, "nd_ok_lean_rej": 0, "nd_rej_lean_ok": 1, "both_rej": 831, "parse_reasons": {"unbound": 547, "name": 93, "expected (": 87, "no-eos": 53, "formula": 47, "ORE box": 28, "expected )": 16, "trailing tokens": 9, "expected ;": 8, "exact not last": 6, "expected :=": 1}, "lean_wall_s": 3.1443822383880615, "lean_proc_s": 7.771884918212891, "nd_verify_s": 0.04659700393676758, "workers": 96, "chunk": 400}
|
| 45 |
+
{"utc": "2026-10-08T00:03:10Z", "samples": 1872, "parse_fail": 959, "distinct_checked": 874, "both_ok": 148, "nd_ok_lean_rej": 0, "nd_rej_lean_ok": 0, "both_rej": 726, "parse_reasons": {"unbound": 646, "expected (": 90, "formula": 45, "no-eos": 45, "name": 41, "ORE box": 41, "expected :=": 24, "expected )": 13, "expected ;": 8, "exact not last": 5, "PR position": 1}, "lean_wall_s": 3.6889543533325195, "lean_proc_s": 8.57895803451538, "nd_verify_s": 0.049927473068237305, "workers": 96, "chunk": 400}
|
| 46 |
+
{"utc": "2026-10-08T00:04:49Z", "samples": 12000, "parse_fail": 1444, "distinct_checked": 10332, "both_ok": 7857, "nd_ok_lean_rej": 0, "nd_rej_lean_ok": 0, "both_rej": 2475, "parse_reasons": {"unbound": 1020, "name": 104, "formula": 88, "expected (": 85, "exact not last": 39, "expected )": 36, "no-eos": 30, "ORE box": 18, "expected ;": 13, "trailing tokens": 5, "expected None": 4, "expected :=": 1, "PR position": 1}, "lean_wall_s": 28.98722267150879, "lean_proc_s": 740.1215221881866, "nd_verify_s": 0.6784279346466064, "workers": 96, "chunk": 400}
|
| 47 |
+
{"utc": "2026-10-08T00:05:03Z", "samples": 1896, "parse_fail": 831, "distinct_checked": 994, "both_ok": 77, "nd_ok_lean_rej": 0, "nd_rej_lean_ok": 0, "both_rej": 917, "parse_reasons": {"unbound": 545, "expected (": 101, "name": 62, "no-eos": 35, "ORE box": 34, "expected )": 24, "formula": 21, "exact not last": 5, "expected ;": 3, "expected fun": 1}, "lean_wall_s": 4.053293228149414, "lean_proc_s": 10.613597869873047, "nd_verify_s": 0.04814553260803223, "workers": 96, "chunk": 400}
|
| 48 |
+
{"utc": "2026-10-08T00:05:20Z", "samples": 1872, "parse_fail": 805, "distinct_checked": 991, "both_ok": 142, "nd_ok_lean_rej": 0, "nd_rej_lean_ok": 0, "both_rej": 849, "parse_reasons": {"unbound": 554, "expected (": 114, "name": 53, "ORE box": 28, "no-eos": 28, "expected )": 13, "formula": 10, "exact not last": 3, "PR position": 1, "expected :=": 1}, "lean_wall_s": 4.0281219482421875, "lean_proc_s": 10.281508684158325, "nd_verify_s": 0.055602073669433594, "workers": 96, "chunk": 400}
|
| 49 |
+
{"utc": "2026-10-08T00:05:44Z", "samples": 1872, "parse_fail": 803, "distinct_checked": 998, "both_ok": 244, "nd_ok_lean_rej": 0, "nd_rej_lean_ok": 0, "both_rej": 754, "parse_reasons": {"unbound": 590, "expected (": 88, "name": 50, "expected )": 28, "ORE box": 24, "formula": 12, "exact not last": 5, "no-eos": 2, "expected ;": 2, "expected None": 1, "PR position": 1}, "lean_wall_s": 3.5331332683563232, "lean_proc_s": 9.559144258499146, "nd_verify_s": 0.0510861873626709, "workers": 96, "chunk": 400}
|
| 50 |
+
{"utc": "2026-10-08T00:07:29Z", "samples": 12000, "parse_fail": 1658, "distinct_checked": 10160, "both_ok": 7806, "nd_ok_lean_rej": 0, "nd_rej_lean_ok": 1, "both_rej": 2353, "parse_reasons": {"unbound": 1165, "no-eos": 95, "name": 76, "formula": 75, "expected (": 67, "exact not last": 48, "expected )": 29, "trailing tokens": 27, "expected ;": 26, "expected :=": 24, "ORE box": 21, "expected None": 3, "PR position": 1, "expected =>": 1}, "lean_wall_s": 29.547099113464355, "lean_proc_s": 741.8609838485718, "nd_verify_s": 0.6664690971374512, "workers": 96, "chunk": 400}
|
| 51 |
+
{"utc": "2026-10-08T00:07:42Z", "samples": 1800, "parse_fail": 879, "distinct_checked": 882, "both_ok": 72, "nd_ok_lean_rej": 0, "nd_rej_lean_ok": 0, "both_rej": 810, "parse_reasons": {"unbound": 597, "expected (": 66, "name": 63, "ORE box": 32, "no-eos": 32, "formula": 30, "expected )": 23, "expected :=": 20, "exact not last": 6, "expected ;": 5, "trailing tokens": 4, "expected None": 1}, "lean_wall_s": 3.8567590713500977, "lean_proc_s": 8.970393419265747, "nd_verify_s": 0.06438422203063965, "workers": 96, "chunk": 400}
|
| 52 |
+
{"utc": "2026-10-08T00:07:58Z", "samples": 1800, "parse_fail": 746, "distinct_checked": 1000, "both_ok": 216, "nd_ok_lean_rej": 0, "nd_rej_lean_ok": 0, "both_rej": 784, "parse_reasons": {"unbound": 558, "expected (": 64, "name": 36, "formula": 28, "ORE box": 27, "expected )": 14, "no-eos": 10, "exact not last": 3, "expected ;": 2, "eof": 1, "expected None": 1, "expected =>": 1, "PR position": 1}, "lean_wall_s": 3.842512607574463, "lean_proc_s": 9.802761554718018, "nd_verify_s": 0.044660329818725586, "workers": 96, "chunk": 400}
|
| 53 |
+
{"utc": "2026-10-08T00:08:20Z", "samples": 1776, "parse_fail": 863, "distinct_checked": 885, "both_ok": 114, "nd_ok_lean_rej": 0, "nd_rej_lean_ok": 0, "both_rej": 771, "parse_reasons": {"unbound": 655, "expected (": 62, "name": 53, "ORE box": 39, "expected )": 31, "expected ;": 8, "no-eos": 7, "exact not last": 3, "formula": 2, "expected =>": 2, "expected False": 1}, "lean_wall_s": 2.84195613861084, "lean_proc_s": 6.307346820831299, "nd_verify_s": 0.04361367225646973, "workers": 96, "chunk": 400}
|
| 54 |
+
{"utc": "2026-10-08T00:10:05Z", "samples": 12000, "parse_fail": 1544, "distinct_checked": 10279, "both_ok": 7849, "nd_ok_lean_rej": 0, "nd_rej_lean_ok": 0, "both_rej": 2430, "parse_reasons": {"unbound": 1126, "no-eos": 104, "name": 94, "expected (": 53, "formula": 49, "expected )": 48, "exact not last": 43, "expected ;": 11, "ORE box": 8, "expected :=": 4, "PR position": 2, "expected None": 2}, "lean_wall_s": 29.503506422042847, "lean_proc_s": 747.1910881996155, "nd_verify_s": 0.6845109462738037, "workers": 96, "chunk": 400}
|
| 55 |
+
{"utc": "2026-10-08T00:10:18Z", "samples": 1824, "parse_fail": 801, "distinct_checked": 989, "both_ok": 154, "nd_ok_lean_rej": 0, "nd_rej_lean_ok": 0, "both_rej": 835, "parse_reasons": {"unbound": 563, "name": 89, "no-eos": 43, "formula": 25, "expected (": 25, "expected :=": 22, "expected )": 15, "exact not last": 6, "ORE box": 4, "expected ;": 4, "expected =>": 3, "trailing tokens": 1, "expected None": 1}, "lean_wall_s": 4.173968553543091, "lean_proc_s": 10.751060009002686, "nd_verify_s": 0.09145832061767578, "workers": 96, "chunk": 400}
|
| 56 |
+
{"utc": "2026-10-08T00:10:35Z", "samples": 1800, "parse_fail": 847, "distinct_checked": 908, "both_ok": 161, "nd_ok_lean_rej": 0, "nd_rej_lean_ok": 0, "both_rej": 747, "parse_reasons": {"unbound": 600, "no-eos": 92, "name": 68, "expected (": 28, "ORE box": 25, "expected )": 22, "exact not last": 3, "formula": 3, "trailing tokens": 2, "expected ;": 2, "expected :=": 2}, "lean_wall_s": 3.998448371887207, "lean_proc_s": 9.966222524642944, "nd_verify_s": 0.04688835144042969, "workers": 96, "chunk": 400}
|
| 57 |
+
{"utc": "2026-10-08T00:11:00Z", "samples": 1800, "parse_fail": 779, "distinct_checked": 979, "both_ok": 140, "nd_ok_lean_rej": 0, "nd_rej_lean_ok": 0, "both_rej": 839, "parse_reasons": {"unbound": 552, "name": 67, "expected (": 63, "no-eos": 58, "formula": 20, "expected )": 15, "exact not last": 2, "expected ;": 1, "expected None": 1}, "lean_wall_s": 4.212545871734619, "lean_proc_s": 10.769506692886353, "nd_verify_s": 0.05795168876647949, "workers": 96, "chunk": 400}
|
| 58 |
+
{"utc": "2026-10-08T00:25:18Z", "samples": 70912, "parse_fail": 14715, "distinct_checked": 49857, "both_ok": 29631, "nd_ok_lean_rej": 0, "nd_rej_lean_ok": 1, "both_rej": 20225, "parse_reasons": {"unbound": 11050, "name": 1264, "expected (": 860, "ORE box": 394, "no-eos": 372, "formula": 243, "expected )": 232, "expected ;": 176, "exact not last": 93, "expected :=": 14, "PR position": 7, "expected None": 5, "trailing tokens": 4, "eof": 1}, "lean_wall_s": 129.46130871772766, "lean_proc_s": 10219.87434887886, "nd_verify_s": 3.1297879219055176, "workers": 96, "chunk": 400}
|
| 59 |
+
{"utc": "2026-10-08T00:25:33Z", "samples": 2000, "parse_fail": 2, "distinct_checked": 1998, "both_ok": 1971, "nd_ok_lean_rej": 0, "nd_rej_lean_ok": 0, "both_rej": 27, "parse_reasons": {"ORE box": 1, "unbound": 1}, "lean_wall_s": 3.556011199951172, "lean_proc_s": 16.51028800010681, "nd_verify_s": 0.10979270935058594, "workers": 96, "chunk": 400}
|
| 60 |
+
{"utc": "2026-10-08T00:31:54Z", "samples": 70912, "parse_fail": 15440, "distinct_checked": 49447, "both_ok": 27344, "nd_ok_lean_rej": 0, "nd_rej_lean_ok": 0, "both_rej": 22103, "parse_reasons": {"unbound": 11428, "name": 1174, "expected (": 777, "no-eos": 568, "ORE box": 476, "formula": 399, "expected )": 279, "expected ;": 213, "exact not last": 99, "expected :=": 11, "PR position": 9, "expected None": 5, "trailing tokens": 2}, "lean_wall_s": 126.79229092597961, "lean_proc_s": 10010.14809346199, "nd_verify_s": 3.0517466068267822, "workers": 96, "chunk": 400}
|
runs/510-fable_more-shorter-rounds_s1/pretrain.pt
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:4b3e23c45293cd48ef8fe9dd319b1720d73bf09ef01616a713ae2b0838695d79
|
| 3 |
+
size 25625153
|
runs/510-fable_more-shorter-rounds_s1/source/harness.py
ADDED
|
@@ -0,0 +1,529 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
#!/usr/bin/env python3
|
| 2 |
+
# imports the fork (model, tokenizer, train, ...), which is not resolvable here
|
| 3 |
+
# pyright: reportMissingImports=false
|
| 4 |
+
"""FROZEN harness for the test-time-loops autoresearch loop. The research agent does not edit it.
|
| 5 |
+
|
| 6 |
+
Forked from ../autoresearch/harness.py (sha 2aa52f23e7ec) on 2026-10-07. What changed: the RL
|
| 7 |
+
stage belongs to the agent, behind a wall-clock budget and a sampling oracle, and the metric is
|
| 8 |
+
measured with the model's test-time compute knob (KNOB) set to EVAL_LEVEL.
|
| 9 |
+
|
| 10 |
+
One timed run = harness owns the clocks, the oracle and the eval; `pretrain.py` (agent-editable)
|
| 11 |
+
owns pretraining and RL:
|
| 12 |
+
|
| 13 |
+
1. load frozen data (Dan's cap-6 Stage-1 file, ladder RL-target pool, ladder transfer pool)
|
| 14 |
+
2. call pretrain.pretrain(train_recs, PRETRAIN_SECS, ...) -> (model, tok, stats); charged for
|
| 15 |
+
everything including model construction and torch.compile. Overrun past the grace -> timeout.
|
| 16 |
+
3. check n_params <= PARAM_CAP, tok.mode == TOK_MODE and that the model has the KNOB attribute
|
| 17 |
+
4. call pretrain.rl(model, tok, targets, train_recs, oracle, RL_SECS, ...): expert iteration in
|
| 18 |
+
whatever form the agent likes. The targets are statements only, and `oracle.sample` is the
|
| 19 |
+
only way to turn them into training data: it samples the model's own attempts and returns
|
| 20 |
+
the ones Lean AND nd_verify accept. RL_SECS is wall-clock inside rl() minus the time the
|
| 21 |
+
oracle spent in the Lean gate (pod CPUs differ; GPU work is what is rationed). Overrun past
|
| 22 |
+
the grace, or an oracle call after the budget is spent -> timeout.
|
| 23 |
+
5. eval on the dev half of the transfer pool with the harness setting the model's KNOB:
|
| 24 |
+
METRIC = dev theorems solved at L_true >= 7 with EVAL_K attempts at EVAL_LEVEL.
|
| 25 |
+
The same eval at DIAG_LEVEL is recorded as a diagnostic, not scored.
|
| 26 |
+
6. SIGALRM at TOTAL_SECS from process start kills everything -> status "timeout"
|
| 27 |
+
|
| 28 |
+
The holdout half of the transfer pool is never scored here. Everything is written to --out.
|
| 29 |
+
|
| 30 |
+
python3 harness.py --out artifacts/x # on the pod (see pod.py / remote.sh)
|
| 31 |
+
ND_FORK=... python3 harness.py --out /tmp/smoke --smoke # CPU, ~1 min, plumbing check
|
| 32 |
+
"""
|
| 33 |
+
|
| 34 |
+
import argparse
|
| 35 |
+
import collections
|
| 36 |
+
import hashlib
|
| 37 |
+
import importlib
|
| 38 |
+
import json
|
| 39 |
+
import os
|
| 40 |
+
import random
|
| 41 |
+
import shutil
|
| 42 |
+
import signal
|
| 43 |
+
import sys
|
| 44 |
+
import time
|
| 45 |
+
import traceback
|
| 46 |
+
from pathlib import Path
|
| 47 |
+
|
| 48 |
+
HERE = os.path.dirname(os.path.abspath(__file__))
|
| 49 |
+
FORK = os.environ.get("ND_FORK", "/workspace/nd-takehome")
|
| 50 |
+
sys.path.insert(0, HERE)
|
| 51 |
+
sys.path.insert(
|
| 52 |
+
0, FORK
|
| 53 |
+
) # fork first: its train.py provides batch/loss_on; the agent's file is pretrain.py
|
| 54 |
+
import torch # noqa: E402
|
| 55 |
+
from eval_set import judge, wilson # noqa: E402
|
| 56 |
+
from sample import generate # noqa: E402
|
| 57 |
+
|
| 58 |
+
# ---------------------------------------------------------------- the rules
|
| 59 |
+
# (hash-checked into every result)
|
| 60 |
+
TOTAL_SECS = int(
|
| 61 |
+
os.environ.get("ND_TOTAL_SECS", 5400)
|
| 62 |
+
) # process start -> summary written; SIGALRM past this. env: benchmarks only
|
| 63 |
+
PRETRAIN_SECS = 300 # agent's pretraining budget, compile included
|
| 64 |
+
PRETRAIN_GRACE = 15 # seconds the agent may overrun before the run is a "timeout"
|
| 65 |
+
RL_SECS = 1800 # agent's RL budget: wall-clock inside rl() minus the oracle's Lean-gate time
|
| 66 |
+
RL_GRACE = 60
|
| 67 |
+
PARAM_CAP = 10_000_000 # unique trainable params, embeddings included; checked after RL too
|
| 68 |
+
N_TARGETS, TEMP, MAX_NEW, SAMPLE_BATCH = 1500, 0.8, 512, 1024
|
| 69 |
+
EVAL_K = 64
|
| 70 |
+
# The model's test-time compute knob: an integer attribute the harness sets before each eval and
|
| 71 |
+
# rl() sets before oracle calls. A sibling loop for another architecture changes these three
|
| 72 |
+
# constants and nothing else in this file.
|
| 73 |
+
KNOB = "n_loops" # loops of the shared core
|
| 74 |
+
EVAL_LEVEL = 8 # the metric: the knob's value during the scored eval
|
| 75 |
+
DIAG_LEVEL = 2 # the same eval again at this value, recorded but not scored
|
| 76 |
+
TOK_MODE = "lean_seq" # required tokenizer mode (fork lean_tok.py)
|
| 77 |
+
DEV = "cuda" if torch.cuda.is_available() else "cpu"
|
| 78 |
+
|
| 79 |
+
|
| 80 |
+
def read(fn):
|
| 81 |
+
return [json.loads(line) for line in open(fn) if line.strip()]
|
| 82 |
+
|
| 83 |
+
|
| 84 |
+
def log(msg):
|
| 85 |
+
print(f"[{time.strftime('%H:%M:%S')}] {msg}", flush=True)
|
| 86 |
+
|
| 87 |
+
|
| 88 |
+
def sha(fn):
|
| 89 |
+
return hashlib.sha256(open(fn, "rb").read()).hexdigest()[:12]
|
| 90 |
+
|
| 91 |
+
|
| 92 |
+
class Timeout(Exception):
|
| 93 |
+
pass
|
| 94 |
+
|
| 95 |
+
|
| 96 |
+
class BudgetSpent(Timeout):
|
| 97 |
+
"""rl() asked the oracle for samples after RL_SECS was used up."""
|
| 98 |
+
|
| 99 |
+
|
| 100 |
+
# ---------------------------------------------------------------- bookkeeping
|
| 101 |
+
|
| 102 |
+
|
| 103 |
+
def save_checkpoint(a, model, tok, stage):
|
| 104 |
+
"""Publish complete weight snapshots atomically; these are not optimizer-resume files."""
|
| 105 |
+
path = Path(a.out) / f"{stage}.pt"
|
| 106 |
+
pending = path.with_suffix(".pt.tmp")
|
| 107 |
+
try:
|
| 108 |
+
torch.save(
|
| 109 |
+
{
|
| 110 |
+
"state": model.state_dict(),
|
| 111 |
+
"tok_mode": getattr(tok, "mode", None),
|
| 112 |
+
"stage": stage,
|
| 113 |
+
"seed": a.seed,
|
| 114 |
+
"harness_sha": sha(f"{HERE}/harness.py"),
|
| 115 |
+
"pretrain_sha": sha(f"{HERE}/pretrain.py"),
|
| 116 |
+
},
|
| 117 |
+
pending,
|
| 118 |
+
)
|
| 119 |
+
os.replace(pending, path)
|
| 120 |
+
finally:
|
| 121 |
+
pending.unlink(missing_ok=True)
|
| 122 |
+
|
| 123 |
+
|
| 124 |
+
def lstar(solved_names, recs, need=5):
|
| 125 |
+
ls_solved = [r["n_lines"] for r in recs if r["name"] in solved_names]
|
| 126 |
+
ge = {L: sum(1 for x in ls_solved if x >= L) for L in range(7, 15)}
|
| 127 |
+
return max([L for L, c in ge.items() if c >= need], default=0), ge
|
| 128 |
+
|
| 129 |
+
|
| 130 |
+
def by_bin(solved_names, recs):
|
| 131 |
+
by = collections.defaultdict(lambda: [0, 0])
|
| 132 |
+
for r in recs:
|
| 133 |
+
by[r["n_lines"]][0] += r["name"] in solved_names
|
| 134 |
+
by[r["n_lines"]][1] += 1
|
| 135 |
+
return {
|
| 136 |
+
str(L): {"solved": k, "n": n, "rate": k / n, "ci": list(wilson(k, n))}
|
| 137 |
+
for L, (k, n) in sorted(by.items())
|
| 138 |
+
}
|
| 139 |
+
|
| 140 |
+
|
| 141 |
+
def gate_seconds():
|
| 142 |
+
"""Seconds the fork's lean_gate has logged so far (Lean wall-clock plus nd_verify)."""
|
| 143 |
+
fn = os.environ.get("LEAN_GATE_LOG")
|
| 144 |
+
if not fn or not os.path.exists(fn):
|
| 145 |
+
return 0.0
|
| 146 |
+
return sum(rec["lean_wall_s"] + rec["nd_verify_s"] for rec in read(fn))
|
| 147 |
+
|
| 148 |
+
|
| 149 |
+
def sample_and_judge(model, tok, recs, k, temperature, seed, batch, max_new):
|
| 150 |
+
prompts = [r["prompt"] for r in recs for _ in range(k)]
|
| 151 |
+
t0, g0 = time.time(), gate_seconds()
|
| 152 |
+
flat = generate(
|
| 153 |
+
model,
|
| 154 |
+
tok,
|
| 155 |
+
prompts,
|
| 156 |
+
greedy=(k == 1 and temperature == 0),
|
| 157 |
+
temperature=temperature,
|
| 158 |
+
batch=batch,
|
| 159 |
+
seed=seed,
|
| 160 |
+
max_new=max_new,
|
| 161 |
+
)
|
| 162 |
+
secs, gate = time.time() - t0, gate_seconds() - g0
|
| 163 |
+
rows = judge(recs, [flat[i * k : (i + 1) * k] for i in range(len(recs))], "n_lines")
|
| 164 |
+
return rows, {
|
| 165 |
+
"samples": len(prompts),
|
| 166 |
+
"secs": secs,
|
| 167 |
+
"gate_secs": gate,
|
| 168 |
+
"samples_per_sec": len(prompts) / secs,
|
| 169 |
+
}
|
| 170 |
+
|
| 171 |
+
|
| 172 |
+
def greedy_rate(model, tok, recs):
|
| 173 |
+
rows, _ = sample_and_judge(model, tok, recs, 1, 0, 0, SAMPLE_BATCH, MAX_NEW)
|
| 174 |
+
return sum(r["solved"] for r in rows) / max(len(rows), 1)
|
| 175 |
+
|
| 176 |
+
|
| 177 |
+
def dev_split(transfer):
|
| 178 |
+
return [r for r in transfer if int(hashlib.sha1(r["key"].encode()).hexdigest(), 16) % 2 == 0]
|
| 179 |
+
|
| 180 |
+
|
| 181 |
+
def holdout_split(transfer):
|
| 182 |
+
"""Never scored by the loop. Score final.pt on this offline when checking in."""
|
| 183 |
+
return [r for r in transfer if int(hashlib.sha1(r["key"].encode()).hexdigest(), 16) % 2 == 1]
|
| 184 |
+
|
| 185 |
+
|
| 186 |
+
# ---------------------------------------------------------------- the RL oracle
|
| 187 |
+
|
| 188 |
+
|
| 189 |
+
class Oracle:
|
| 190 |
+
"""What rl() gets instead of data: the model's own attempts at the RL targets, filtered to
|
| 191 |
+
the ones Lean AND nd_verify accept. It also keeps the RL clock and the record of every call.
|
| 192 |
+
"""
|
| 193 |
+
|
| 194 |
+
def __init__(self, a, tok, targets):
|
| 195 |
+
self.a, self.tok = a, tok
|
| 196 |
+
self.targets = {t["name"]: t for t in targets}
|
| 197 |
+
self.found = collections.defaultdict(dict) # name -> {proof text: record}
|
| 198 |
+
self.calls = []
|
| 199 |
+
self.t0 = self.last = time.time()
|
| 200 |
+
self.uncharged = 0.0
|
| 201 |
+
|
| 202 |
+
def used(self):
|
| 203 |
+
"""RL seconds charged so far: wall-clock since rl() started, minus Lean-gate time."""
|
| 204 |
+
return time.time() - self.t0 - self.uncharged
|
| 205 |
+
|
| 206 |
+
def left(self):
|
| 207 |
+
return self.a.rl_secs - self.used()
|
| 208 |
+
|
| 209 |
+
def sample(self, model, targets, k, temperature=TEMP, batch=SAMPLE_BATCH):
|
| 210 |
+
"""k attempts at each of `targets` (records rl() was given, any subset, in any order) from
|
| 211 |
+
`model` as it is now, in eval mode, at whatever its KNOB is set to. Returns one
|
| 212 |
+
dict per target, in order: name, prompt, n_lines, proofs (the distinct accepted proof
|
| 213 |
+
texts), n_ok (accepted attempts) and n_tried. Rejected attempts are not returned.
|
| 214 |
+
Raises BudgetSpent, which ends the run as a timeout, once the budget is used up."""
|
| 215 |
+
if self.left() <= 0:
|
| 216 |
+
raise BudgetSpent(f"oracle called with {self.used():.0f}s of {self.a.rl_secs}s used")
|
| 217 |
+
recs = [self.targets[t["name"]] for t in targets]
|
| 218 |
+
call = len(self.calls) + 1
|
| 219 |
+
model.eval()
|
| 220 |
+
rows, samp = sample_and_judge(
|
| 221 |
+
model, self.tok, recs, k, temperature, self.a.seed * 1000 + call, batch, self.a.max_new
|
| 222 |
+
)
|
| 223 |
+
self.uncharged += samp["gate_secs"]
|
| 224 |
+
new = 0
|
| 225 |
+
for t, row in zip(recs, rows, strict=True):
|
| 226 |
+
for p, wl, pl in zip(
|
| 227 |
+
row["proofs"], row["written_lens"], row["pruned_lens"], strict=True
|
| 228 |
+
):
|
| 229 |
+
if p not in self.found[t["name"]]:
|
| 230 |
+
self.found[t["name"]][p] = {
|
| 231 |
+
"proof": p,
|
| 232 |
+
"written": wl,
|
| 233 |
+
"pruned": pl,
|
| 234 |
+
"round": call,
|
| 235 |
+
}
|
| 236 |
+
new += 1
|
| 237 |
+
solved = {n for n, d in self.found.items() if d}
|
| 238 |
+
ls, ge = lstar(solved, list(self.targets.values()))
|
| 239 |
+
now = time.time()
|
| 240 |
+
self.calls.append(
|
| 241 |
+
{
|
| 242 |
+
"round": call,
|
| 243 |
+
"level": getattr(model, KNOB, None),
|
| 244 |
+
"k": k,
|
| 245 |
+
"n_targets": len(recs),
|
| 246 |
+
"sample_acc": sum(x["n_ok"] for x in rows) / max(samp["samples"], 1),
|
| 247 |
+
"new_proofs": new,
|
| 248 |
+
"targets_solved": len(solved),
|
| 249 |
+
"lstar": ls,
|
| 250 |
+
"ge": ge,
|
| 251 |
+
"mix_rl_records": sum(len(x["proofs"]) for x in rows), # proofs handed back
|
| 252 |
+
"sampling": samp,
|
| 253 |
+
"secs": now - self.last, # since the previous call returned: includes rl()'s work
|
| 254 |
+
"rl_secs_used": self.used(),
|
| 255 |
+
}
|
| 256 |
+
)
|
| 257 |
+
self.last = now
|
| 258 |
+
c = self.calls[-1]
|
| 259 |
+
log(
|
| 260 |
+
f"ei r{call}: acc {c['sample_acc']:.4f}, +{new} proofs, "
|
| 261 |
+
f"solved {len(solved)}/{len(self.targets)}, L*={ls}, {KNOB} {c['level']}, "
|
| 262 |
+
f"{len(recs)} targets x {k}, {samp['samples_per_sec']:.0f} samples/s, "
|
| 263 |
+
f"RL clock {c['rl_secs_used']:.0f}/{self.a.rl_secs}s"
|
| 264 |
+
)
|
| 265 |
+
return [
|
| 266 |
+
{
|
| 267 |
+
"name": t["name"],
|
| 268 |
+
"prompt": t["prompt"],
|
| 269 |
+
"n_lines": t["n_lines"],
|
| 270 |
+
"proofs": list(row["proofs"]),
|
| 271 |
+
"n_ok": row["n_ok"],
|
| 272 |
+
"n_tried": row["n_tried"],
|
| 273 |
+
}
|
| 274 |
+
for t, row in zip(recs, rows, strict=True)
|
| 275 |
+
]
|
| 276 |
+
|
| 277 |
+
|
| 278 |
+
# ---------------------------------------------------------------- stages
|
| 279 |
+
|
| 280 |
+
|
| 281 |
+
def stage_rl(a, agent, tok, model, targets, train_recs):
|
| 282 |
+
oracle = Oracle(a, tok, targets)
|
| 283 |
+
# statements only: what the agent's RL may know about a target
|
| 284 |
+
visible = [{"name": t["name"], "prompt": t["prompt"], "n_lines": t["n_lines"]} for t in targets]
|
| 285 |
+
stats = agent.rl(model, tok, visible, train_recs, oracle, a.rl_secs, DEV, a.seed, log)
|
| 286 |
+
if DEV == "cuda":
|
| 287 |
+
torch.cuda.synchronize()
|
| 288 |
+
used = oracle.used()
|
| 289 |
+
model.eval()
|
| 290 |
+
with open(f"{a.out}/found_targets.jsonl", "w") as f:
|
| 291 |
+
for t in targets:
|
| 292 |
+
for x in oracle.found[t["name"]].values():
|
| 293 |
+
f.write(
|
| 294 |
+
json.dumps({"name": t["name"], "thm": t["thm"], "L_true": t["n_lines"], **x})
|
| 295 |
+
+ "\n"
|
| 296 |
+
)
|
| 297 |
+
log(f"rl returned: {used:.0f}s of {a.rl_secs}s charged, {oracle.uncharged:.0f}s Lean gate")
|
| 298 |
+
return {
|
| 299 |
+
"rounds": oracle.calls,
|
| 300 |
+
"secs": time.time() - oracle.t0,
|
| 301 |
+
"rl_secs_used": used,
|
| 302 |
+
"gate_secs": oracle.uncharged,
|
| 303 |
+
"targets": len(targets),
|
| 304 |
+
"stats": dict(stats or {}),
|
| 305 |
+
}
|
| 306 |
+
|
| 307 |
+
|
| 308 |
+
def eval_at(a, tok, model, dev_pool, level):
|
| 309 |
+
setattr(model, KNOB, level)
|
| 310 |
+
rows, samp = sample_and_judge(
|
| 311 |
+
model, tok, dev_pool, a.eval_k, TEMP, a.seed * 1000 + 777, SAMPLE_BATCH, a.max_new
|
| 312 |
+
)
|
| 313 |
+
solved = {r["name"] for r in rows if r["solved"]}
|
| 314 |
+
ls, ge = lstar(solved, dev_pool)
|
| 315 |
+
out = {
|
| 316 |
+
"level": level,
|
| 317 |
+
"attempts_per_theorem": a.eval_k,
|
| 318 |
+
"solved": len(solved),
|
| 319 |
+
"n": len(dev_pool),
|
| 320 |
+
"solved_ge7": ge[7],
|
| 321 |
+
"lstar": ls,
|
| 322 |
+
"ge": ge,
|
| 323 |
+
"by_bin": by_bin(solved, dev_pool),
|
| 324 |
+
"sample_acc": sum(r["n_ok"] for r in rows) / max(samp["samples"], 1),
|
| 325 |
+
"max_pruned": max((max(r["pruned_lens"]) for r in rows if r["solved"]), default=0),
|
| 326 |
+
"sampling": samp,
|
| 327 |
+
}
|
| 328 |
+
log(
|
| 329 |
+
f"eval at {KNOB} {level}: solved {len(solved)}/{len(dev_pool)}, ge7 {ge[7]}, L*={ls}, "
|
| 330 |
+
f"per-sample {out['sample_acc']:.4f}, {samp['samples_per_sec']:.0f} samples/s"
|
| 331 |
+
)
|
| 332 |
+
return out, rows
|
| 333 |
+
|
| 334 |
+
|
| 335 |
+
def stage_eval(a, tok, model, dev_pool, heldout_recs):
|
| 336 |
+
out, rows = eval_at(a, tok, model, dev_pool, a.eval_level)
|
| 337 |
+
out["heldout_greedy"] = greedy_rate(model, tok, heldout_recs)
|
| 338 |
+
with open(f"{a.out}/eval_dev.jsonl", "w") as f:
|
| 339 |
+
for r in rows:
|
| 340 |
+
keys = ("name", "n_lines", "solved", "n_ok", "proofs", "written_lens", "pruned_lens")
|
| 341 |
+
f.write(json.dumps({k: r[k] for k in keys}) + "\n")
|
| 342 |
+
diag, _ = eval_at(a, tok, model, dev_pool, a.diag_level)
|
| 343 |
+
out["diag"] = diag
|
| 344 |
+
setattr(model, KNOB, a.eval_level)
|
| 345 |
+
log(
|
| 346 |
+
f"eval: solved {out['solved']}/{len(dev_pool)}, ge7 {out['solved_ge7']} at "
|
| 347 |
+
f"{KNOB} {a.eval_level} ({diag['solved_ge7']} at {a.diag_level}), L*={out['lstar']}, "
|
| 348 |
+
f"max pruned {out['max_pruned']}, held-out greedy {out['heldout_greedy']:.3f}"
|
| 349 |
+
)
|
| 350 |
+
return out
|
| 351 |
+
|
| 352 |
+
|
| 353 |
+
def lean_gate_totals():
|
| 354 |
+
"""Sum the fork lean_gate's per-generate() log lines (remote.sh sets LEAN_GATE_LOG)."""
|
| 355 |
+
fn = os.environ.get("LEAN_GATE_LOG")
|
| 356 |
+
if not fn or not os.path.exists(fn):
|
| 357 |
+
return None
|
| 358 |
+
keys = ("samples", "parse_fail", "distinct_checked", "both_ok", "nd_ok_lean_rej")
|
| 359 |
+
tot = dict.fromkeys(keys + ("nd_rej_lean_ok", "lean_wall_s"), 0)
|
| 360 |
+
for rec in read(fn):
|
| 361 |
+
for k in tot:
|
| 362 |
+
tot[k] += rec[k]
|
| 363 |
+
return tot
|
| 364 |
+
|
| 365 |
+
|
| 366 |
+
RULES = (
|
| 367 |
+
"TOTAL_SECS",
|
| 368 |
+
"PRETRAIN_SECS",
|
| 369 |
+
"RL_SECS",
|
| 370 |
+
"PARAM_CAP",
|
| 371 |
+
"N_TARGETS",
|
| 372 |
+
"EVAL_K",
|
| 373 |
+
"KNOB",
|
| 374 |
+
"EVAL_LEVEL",
|
| 375 |
+
"DIAG_LEVEL",
|
| 376 |
+
"TOK_MODE",
|
| 377 |
+
)
|
| 378 |
+
|
| 379 |
+
|
| 380 |
+
def write_summary(a, status, **kw):
|
| 381 |
+
s = {
|
| 382 |
+
"status": status,
|
| 383 |
+
"metric": kw.get("eval", {}).get("solved_ge7"),
|
| 384 |
+
"metric_diag": kw.get("eval", {}).get("diag", {}).get("solved_ge7"),
|
| 385 |
+
"metric_rl": (kw.get("rl", {}).get("rounds") or [{}])[-1].get(
|
| 386 |
+
"targets_solved"
|
| 387 |
+
), # secondary: cumulative RL targets solved
|
| 388 |
+
"harness_sha": sha(f"{HERE}/harness.py"),
|
| 389 |
+
"pretrain_sha": sha(f"{HERE}/pretrain.py"),
|
| 390 |
+
"device": DEV,
|
| 391 |
+
"gpu": torch.cuda.get_device_name(0) if DEV == "cuda" else "cpu",
|
| 392 |
+
"torch": torch.__version__,
|
| 393 |
+
"lean_gate": lean_gate_totals(),
|
| 394 |
+
"total_secs": time.time() - a.T0,
|
| 395 |
+
"rules": {k: globals()[k] for k in RULES},
|
| 396 |
+
**kw,
|
| 397 |
+
}
|
| 398 |
+
with open(f"{a.out}/summary.json", "w") as f:
|
| 399 |
+
json.dump(s, f, indent=1)
|
| 400 |
+
log(
|
| 401 |
+
f"SUMMARY status={status} metric={s['metric']} metric_diag={s['metric_diag']} "
|
| 402 |
+
f"metric_rl={s['metric_rl']} total {s['total_secs']:.0f}s"
|
| 403 |
+
)
|
| 404 |
+
return s
|
| 405 |
+
|
| 406 |
+
|
| 407 |
+
def main():
|
| 408 |
+
ap = argparse.ArgumentParser()
|
| 409 |
+
ap.add_argument("--out", required=True)
|
| 410 |
+
ap.add_argument("--seed", type=int, default=0)
|
| 411 |
+
ap.add_argument(
|
| 412 |
+
"--pretrain-only", action="store_true", help="save pretrain.pt and stop before eval/RL"
|
| 413 |
+
)
|
| 414 |
+
ap.add_argument(
|
| 415 |
+
"--smoke", action="store_true", help="tiny everything, for a CPU plumbing check"
|
| 416 |
+
)
|
| 417 |
+
a = ap.parse_args()
|
| 418 |
+
a.pretrain_secs, a.rl_secs, a.eval_k, a.max_new, a.n_targets = (
|
| 419 |
+
PRETRAIN_SECS,
|
| 420 |
+
RL_SECS,
|
| 421 |
+
EVAL_K,
|
| 422 |
+
MAX_NEW,
|
| 423 |
+
N_TARGETS,
|
| 424 |
+
)
|
| 425 |
+
a.eval_level, a.diag_level = EVAL_LEVEL, DIAG_LEVEL
|
| 426 |
+
if a.smoke:
|
| 427 |
+
a.pretrain_secs, a.rl_secs, a.eval_k, a.max_new, a.n_targets = 10, 30, 2, 120, 8
|
| 428 |
+
os.makedirs(a.out, exist_ok=True)
|
| 429 |
+
# Preserve discarded variants too: weights alone do not specify the model architecture.
|
| 430 |
+
source_dir = Path(a.out) / "source"
|
| 431 |
+
source_dir.mkdir(exist_ok=True)
|
| 432 |
+
for filename in ("pretrain.py", "harness.py", "remote.sh"):
|
| 433 |
+
shutil.copy2(Path(HERE) / filename, source_dir / filename)
|
| 434 |
+
a.T0 = time.time()
|
| 435 |
+
torch.manual_seed(a.seed)
|
| 436 |
+
|
| 437 |
+
def on_alarm(signum, frame):
|
| 438 |
+
raise Timeout()
|
| 439 |
+
|
| 440 |
+
signal.signal(signal.SIGALRM, on_alarm)
|
| 441 |
+
signal.alarm(TOTAL_SECS if not a.smoke else 900)
|
| 442 |
+
|
| 443 |
+
train_recs = read(f"{FORK}/data/train.jsonl")
|
| 444 |
+
assert all(r["n_lines"] <= 6 for r in train_recs), "pretrain file violates cap 6"
|
| 445 |
+
heldout_all = read(f"{FORK}/data/heldout.jsonl")
|
| 446 |
+
# Length-ordered file: take 2000 evenly spaced records, so every length 2-6 is covered.
|
| 447 |
+
heldout_all = [heldout_all[i * len(heldout_all) // 2000] for i in range(2000)]
|
| 448 |
+
heldout_recs = heldout_all if not a.smoke else heldout_all[:16]
|
| 449 |
+
targets_all = read(f"{FORK}/data/ladder/rl_targets.jsonl")
|
| 450 |
+
transfer = read(f"{FORK}/data/ladder/transfer.jsonl")
|
| 451 |
+
dev_pool = dev_split(transfer) if not a.smoke else dev_split(transfer)[:8]
|
| 452 |
+
targets = random.Random(a.seed).sample(targets_all, min(a.n_targets, len(targets_all)))
|
| 453 |
+
tk, dk, hk = (
|
| 454 |
+
{t["key"] for t in targets},
|
| 455 |
+
{r["key"] for r in dev_pool},
|
| 456 |
+
{r["key"] for r in train_recs},
|
| 457 |
+
)
|
| 458 |
+
assert not (tk & dk) and not (tk & hk) and not (dk & hk), "renaming-class leak between pools"
|
| 459 |
+
log(
|
| 460 |
+
f"pools: train {len(train_recs):,}, targets {len(targets)}, dev-transfer {len(dev_pool)}, "
|
| 461 |
+
f"held-out {len(heldout_recs)}; "
|
| 462 |
+
f"device {DEV}; harness {sha(f'{HERE}/harness.py')} pretrain {sha(f'{HERE}/pretrain.py')}"
|
| 463 |
+
)
|
| 464 |
+
|
| 465 |
+
pre = {}
|
| 466 |
+
try:
|
| 467 |
+
agent = importlib.import_module("pretrain")
|
| 468 |
+
t0 = time.time()
|
| 469 |
+
model, tok, pre = agent.pretrain(train_recs, a.pretrain_secs, DEV, a.seed, log)
|
| 470 |
+
if DEV == "cuda":
|
| 471 |
+
torch.cuda.synchronize()
|
| 472 |
+
pre = dict(pre or {})
|
| 473 |
+
pre["secs"] = time.time() - t0
|
| 474 |
+
model.eval()
|
| 475 |
+
n_params = sum(p.numel() for p in model.parameters())
|
| 476 |
+
pre["n_params"] = n_params
|
| 477 |
+
log(
|
| 478 |
+
f"pretrain returned after {pre['secs']:.0f}s (budget {a.pretrain_secs}s), "
|
| 479 |
+
f"{n_params:,} params"
|
| 480 |
+
)
|
| 481 |
+
if pre["secs"] > a.pretrain_secs + PRETRAIN_GRACE and not a.smoke:
|
| 482 |
+
over = pre["secs"] - a.pretrain_secs
|
| 483 |
+
write_summary(
|
| 484 |
+
a, "timeout", pretrain=pre, reason=f"pretrain overran budget by {over:.0f}s"
|
| 485 |
+
)
|
| 486 |
+
return
|
| 487 |
+
if getattr(tok, "mode", None) != TOK_MODE:
|
| 488 |
+
write_summary(a, "crash", pretrain=pre, reason=f"tok.mode {tok.mode!r} != {TOK_MODE!r}")
|
| 489 |
+
return
|
| 490 |
+
if n_params > PARAM_CAP:
|
| 491 |
+
write_summary(a, "crash", pretrain=pre, reason=f"{n_params} params > cap {PARAM_CAP}")
|
| 492 |
+
return
|
| 493 |
+
if not isinstance(getattr(model, KNOB, None), int):
|
| 494 |
+
reason = f"model has no integer {KNOB} for the harness to set at eval"
|
| 495 |
+
write_summary(a, "crash", pretrain=pre, reason=reason)
|
| 496 |
+
return
|
| 497 |
+
save_checkpoint(a, model, tok, "pretrain")
|
| 498 |
+
if a.pretrain_only:
|
| 499 |
+
write_summary(a, "ok", stage="pretrain_only", pretrain=pre)
|
| 500 |
+
return
|
| 501 |
+
pre["heldout_greedy"] = greedy_rate(model, tok, heldout_recs)
|
| 502 |
+
log(f"pretrain held-out greedy {pre['heldout_greedy']:.3f}")
|
| 503 |
+
rl = stage_rl(a, agent, tok, model, targets, train_recs)
|
| 504 |
+
if rl["rl_secs_used"] > a.rl_secs + RL_GRACE and not a.smoke:
|
| 505 |
+
over = rl["rl_secs_used"] - a.rl_secs
|
| 506 |
+
reason = f"rl overran budget by {over:.0f}s"
|
| 507 |
+
write_summary(a, "timeout", pretrain=pre, rl=rl, reason=reason)
|
| 508 |
+
return
|
| 509 |
+
n_after = sum(p.numel() for p in model.parameters())
|
| 510 |
+
if n_after > PARAM_CAP:
|
| 511 |
+
reason = f"{n_after} params > cap {PARAM_CAP} after rl"
|
| 512 |
+
write_summary(a, "crash", pretrain=pre, rl=rl, reason=reason)
|
| 513 |
+
return
|
| 514 |
+
save_checkpoint(a, model, tok, "final")
|
| 515 |
+
ev = stage_eval(a, tok, model, dev_pool, heldout_recs)
|
| 516 |
+
write_summary(a, "ok", pretrain=pre, rl=rl, eval=ev)
|
| 517 |
+
except BudgetSpent as e:
|
| 518 |
+
write_summary(a, "timeout", pretrain=pre, reason=f"rl: {e}")
|
| 519 |
+
except Timeout:
|
| 520 |
+
write_summary(a, "timeout", pretrain=pre, reason=f"exceeded TOTAL_SECS={TOTAL_SECS}")
|
| 521 |
+
except Exception:
|
| 522 |
+
traceback.print_exc()
|
| 523 |
+
write_summary(a, "crash", pretrain=pre, reason=traceback.format_exc()[-2000:])
|
| 524 |
+
finally:
|
| 525 |
+
signal.alarm(0)
|
| 526 |
+
|
| 527 |
+
|
| 528 |
+
if __name__ == "__main__":
|
| 529 |
+
main()
|
runs/510-fable_more-shorter-rounds_s1/source/pretrain.py
ADDED
|
@@ -0,0 +1,611 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# imports the fork (model, tokenizer, train, ...), which is not resolvable here
|
| 2 |
+
# pyright: reportMissingImports=false
|
| 3 |
+
"""AGENT-EDITABLE: pretraining and RL. This is the only file the autoresearch agent changes.
|
| 4 |
+
|
| 5 |
+
Contract with harness.py (frozen):
|
| 6 |
+
|
| 7 |
+
pretrain(train_recs, budget_secs, device, seed, log) -> (model, tok, stats)
|
| 8 |
+
|
| 9 |
+
train_recs list of dicts {prompt, proof, n_lines<=6, key, ...}; the frozen cap-6 Stage-1 data
|
| 10 |
+
budget_secs wall-clock you are charged for, from call to return, INCLUDING model construction
|
| 11 |
+
and torch.compile. Overrun by more than the harness grace -> the run is a timeout.
|
| 12 |
+
model nn.Module; <= PARAM_CAP unique params (checked again after rl). The harness samples
|
| 13 |
+
from it with the fork's sample.generate, so it must provide:
|
| 14 |
+
- forward(idx) -> logits (B,T,V) and forward(idx, pos=, mask=, caches=) with a
|
| 15 |
+
per-block KV cache dict, exactly like the fork's GPT (model.blocks must exist)
|
| 16 |
+
- an integer `n_loops` (the harness's KNOB): the loop count applied in eval
|
| 17 |
+
mode. The harness sets it to EVAL_LEVEL for the scored eval and to DIAG_LEVEL
|
| 18 |
+
for the diagnostic one.
|
| 19 |
+
tok must provide encode_prompt, encode_proof, decode, shift_abs, pad, eos, mode, and
|
| 20 |
+
mode must be "lean_seq" (harness TOK_MODE): Lean text in and out, Lean-gated samples.
|
| 21 |
+
Decoder rule: decode() is deterministic and single-pass; it may not call the
|
| 22 |
+
verifier or try alternatives.
|
| 23 |
+
stats dict recorded in summary.json and copied into results.json. MUST contain `curve`:
|
| 24 |
+
a list of {secs, step, train_loss, val_loss} taken every ~CURVE_EVERY seconds
|
| 25 |
+
(see log_point below), so a run's trajectory survives in git. Add anything else.
|
| 26 |
+
val_loss is on VAL_N records held out from train_recs by this file; keep that.
|
| 27 |
+
|
| 28 |
+
rl(model, tok, targets, train_recs, oracle, budget_secs, device, seed, log) -> stats | None
|
| 29 |
+
|
| 30 |
+
targets list of dicts {name, prompt, n_lines}: the 1,500 RL target statements. No proofs.
|
| 31 |
+
oracle oracle.sample(model, targets_subset, k, temperature=0.8, batch=1024) samples k
|
| 32 |
+
attempts per target from the model in eval mode at its current n_loops, and returns
|
| 33 |
+
per target {name, prompt, n_lines, proofs, n_ok, n_tried}; `proofs` are the distinct
|
| 34 |
+
attempts Lean AND nd_verify accepted. This is the only source of new training data.
|
| 35 |
+
oracle.used() / oracle.left(): seconds of the RL budget charged / remaining.
|
| 36 |
+
budget_secs RL_SECS: wall-clock inside rl() minus the oracle's Lean-gate time. Returning more
|
| 37 |
+
than the grace late, or calling the oracle with nothing left, is a timeout.
|
| 38 |
+
The model is trained in place. Everything else about RL is yours: rounds, samples per target,
|
| 39 |
+
loop counts when sampling, which proofs to keep, weighting, replay, optimizer, loss.
|
| 40 |
+
|
| 41 |
+
Anything imported from the fork (ND_FORK) is fair game to copy in here and modify.
|
| 42 |
+
"""
|
| 43 |
+
|
| 44 |
+
import math
|
| 45 |
+
import random
|
| 46 |
+
import time
|
| 47 |
+
|
| 48 |
+
import torch
|
| 49 |
+
import torch._inductor.config # noqa: F401 (explicit, so the attribute resolves for pyright)
|
| 50 |
+
from lean_tok import MAXN # fork
|
| 51 |
+
from model import GPT, Block # fork
|
| 52 |
+
from normalize import norm # fork
|
| 53 |
+
from tokenizer import make_tokenizer # fork
|
| 54 |
+
from train import loss_on # fork
|
| 55 |
+
|
| 56 |
+
# ---- knobs
|
| 57 |
+
# Unique blocks: one untied entry block, N_CORE shared core blocks, one untied exit block.
|
| 58 |
+
N_CORE = 2
|
| 59 |
+
N_LAYER, D, N_HEAD = 2 + N_CORE, 384, 8
|
| 60 |
+
# Core applications (loops). Training draws one count per batch from TRAIN_LOOPS, in pretraining
|
| 61 |
+
# and in rl()'s fine-tune; eval and sampling apply model.n_loops, which the harness sets for its
|
| 62 |
+
# evals and rl() sets before each oracle call. DEFAULT_LOOPS is used for the validation curve and
|
| 63 |
+
# the harness's post-pretraining greedy check.
|
| 64 |
+
TRAIN_LOOPS = (1, 2, 3, 4, 5, 6, 7, 8)
|
| 65 |
+
DEFAULT_LOOPS = 2
|
| 66 |
+
VAL_LOOPS = (1, 2, 4, 8) # val loss at each count, once, when pretraining ends
|
| 67 |
+
# ---- RL knobs: the old harness's expert iteration (600 AdamW steps per round on up to 4 proofs
|
| 68 |
+
# per solved target, weighted x4, mixed with 20,000 replayed pretraining records), except for
|
| 69 |
+
# where the samples go. Each round every target gets RL_K attempts at RL_FIRST_LOOPS; a target
|
| 70 |
+
# all of them fail on gets RL_RETRY_K more (32 in all, as before) at ONE deeper loop count from
|
| 71 |
+
# RL_RETRY_LOOPS, assigned at random, so proofs that only deeper sampling finds enter the
|
| 72 |
+
# training set and no deep sample is spent on a target one loop already solves. The random
|
| 73 |
+
# assignment makes the retry calls a controlled comparison of loop counts on resistant targets.
|
| 74 |
+
# Rounds repeat until another one would not fit in the budget.
|
| 75 |
+
RL_FIRST_LOOPS, RL_K = 1, 8
|
| 76 |
+
RL_RETRY_LOOPS, RL_RETRY_K = (2, 4, 8), 24
|
| 77 |
+
# The fine-tune sees FT_RECORDS records per round, half the old harness's 600 steps x 128: in
|
| 78 |
+
# 509 the fine-tune had grown to ~74 s of a ~137 s round, so halving it fits about 16 rounds of
|
| 79 |
+
# sampling in RL_SECS instead of 11, with 27% fewer fine-tune records per run; the mix (20-33k
|
| 80 |
+
# records) still gets at least one epoch per round. Records come in the same token-budget
|
| 81 |
+
# batches as pretraining (TOK_BUDGET below) instead of 128-record batches padded to their
|
| 82 |
+
# longest proof, which with 7-14 line RL proofs in nearly every batch was ~3x the content.
|
| 83 |
+
# It uses pretraining's optimizer split: Muon (muon_step, MUON_WD) on the block matrices and the
|
| 84 |
+
# head, AdamW at FT_LR on embeddings, norms and biases. FT_MUON_LR keeps pretraining's 10:1
|
| 85 |
+
# Muon:AdamW ratio and is where pretraining's cosine ends (MUON_LR * MIN_LR_FRAC).
|
| 86 |
+
FT_RECORDS, FT_LR, FT_MUON_LR, FT_WARMUP = 300 * 128, 3e-4, 3e-3, 50
|
| 87 |
+
RETAIN, MAX_PER_THM, RL_WEIGHT = 20000, 4, 4
|
| 88 |
+
# MLP width held at 151's 1280: the fork's Block hardcodes 4 * D (1536 at D 384 would pass the
|
| 89 |
+
# 10M cap), so PeriLNBlock swaps fc1/fc2. 1280 != 3 * D keeps muon_step's QKV-by-shape test valid.
|
| 90 |
+
D_FF = 1280
|
| 91 |
+
BS, LR, MIN_LR_FRAC, WD = 128, 1e-3, 0.3, 0.1
|
| 92 |
+
MUON_LR, MUON_MOMENTUM = 0.01, 0.95
|
| 93 |
+
# Polar Express (Amsel et al. 2025) per-iteration quintic coefficients, as in modded-nanogpt
|
| 94 |
+
# (num_iters=5, safety factor 2e-2): singular values land in ~[0.86, 1.14] instead of the
|
| 95 |
+
# fixed quintic's ~[0.68, 1.13], at the same five iterations.
|
| 96 |
+
NS_COEFFS = (
|
| 97 |
+
(8.156554524902461, -22.48329292557795, 15.878769915207462),
|
| 98 |
+
(4.042929935166739, -2.808917465908714, 0.5000178451051316),
|
| 99 |
+
(3.8916678022926607, -2.772484153217685, 0.5060648178503393),
|
| 100 |
+
(3.285753657755655, -2.3681294933425376, 0.46449024233003106),
|
| 101 |
+
(2.3465413258596377, -1.7097828382687081, 0.42323551169305323),
|
| 102 |
+
)
|
| 103 |
+
# Ten times the AdamW group's per-step multiplicative decay (lr * wd) on the Muon matrices.
|
| 104 |
+
MUON_WD = 10 * WD * LR / MUON_LR
|
| 105 |
+
TOK_MODE = "lean_seq" # required by the harness (Lean protocol)
|
| 106 |
+
# Hybrid heads: half the heads get ALiBi (Press et al. 2021) with the standard slope set for
|
| 107 |
+
# H/2 heads, 2^(-8 i / (H/2)) = 1/4 .. 1/256 with 8 heads; the other half get slope 0, i.e. no
|
| 108 |
+
# positional signal at all (NoPE), so they can look citations up by content at any distance.
|
| 109 |
+
N_ALIBI = N_HEAD // 2
|
| 110 |
+
ALIBI_SLOPES = tuple(2.0 ** (-8.0 * (i + 1) / N_ALIBI) for i in range(N_ALIBI)) + (0.0,) * (
|
| 111 |
+
N_HEAD - N_ALIBI
|
| 112 |
+
)
|
| 113 |
+
VAL_N, CURVE_EVERY = 1024, 15 # records held out for val_loss; seconds between curve points
|
| 114 |
+
# Token-budget batches (fairseq max-tokens): each step takes a run of length-sorted records with
|
| 115 |
+
# count * padded_length <= TOK_BUDGET, so a batch of short proofs has many records and a batch of
|
| 116 |
+
# long ones few, and the padded token count per step is ~constant. 128 x 144 matches the real
|
| 117 |
+
# (unpadded) tokens of the incumbent's 128-record batches (mean record ~135 tokens), whose
|
| 118 |
+
# padding to the batch max (~400) made every step ~3x larger than its content.
|
| 119 |
+
TOK_BUDGET = 128 * 144
|
| 120 |
+
# Multi-token prediction (DeepSeek-V3 2024, Gloeckle et al. 2024), pretraining only: one extra
|
| 121 |
+
# Peri-LN block reads [trunk state at t; embedding of token t+1] and predicts token t+2 through
|
| 122 |
+
# the model's own ln_f/head. Its loss is added at this weight (DeepSeek-V3's early-training value).
|
| 123 |
+
MTP_WEIGHT = 0.3
|
| 124 |
+
|
| 125 |
+
|
| 126 |
+
# Muon update adapted from KellerJordan/Muon, commit
|
| 127 |
+
# f98f1cacc0263b04290753e32be8d498c1efc806 (muon.py).
|
| 128 |
+
# MIT License
|
| 129 |
+
# Copyright (c) 2024 Keller Jordan
|
| 130 |
+
#
|
| 131 |
+
# Permission is hereby granted, free of charge, to any person obtaining a copy
|
| 132 |
+
# of this software and associated documentation files (the "Software"), to deal
|
| 133 |
+
# in the Software without restriction, including without limitation the rights
|
| 134 |
+
# to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
|
| 135 |
+
# copies of the Software, and to permit persons to whom the Software is
|
| 136 |
+
# furnished to do so, subject to the following conditions:
|
| 137 |
+
# The above copyright notice and this permission notice shall be included in all
|
| 138 |
+
# copies or substantial portions of the Software.
|
| 139 |
+
# THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
|
| 140 |
+
# IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
|
| 141 |
+
# FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
|
| 142 |
+
# AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
|
| 143 |
+
# LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
|
| 144 |
+
# OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
|
| 145 |
+
# SOFTWARE.
|
| 146 |
+
@torch.no_grad()
|
| 147 |
+
def muon_step(params, buffers, lr):
|
| 148 |
+
"""Nesterov momentum plus five bf16 Polar Express steps, on hidden matrices only."""
|
| 149 |
+
for p, momentum in zip(params, buffers, strict=True):
|
| 150 |
+
if p.grad is None:
|
| 151 |
+
continue
|
| 152 |
+
momentum.lerp_(p.grad, 1 - MUON_MOMENTUM)
|
| 153 |
+
update = p.grad.lerp(momentum, MUON_MOMENTUM).bfloat16()
|
| 154 |
+
# Only QKV has shape (3D, D) in this GPT; orthogonalize its roles separately.
|
| 155 |
+
if update.size(0) == 3 * update.size(1):
|
| 156 |
+
update = update.reshape(3, update.size(1), update.size(1))
|
| 157 |
+
transpose = update.size(-2) > update.size(-1)
|
| 158 |
+
if transpose:
|
| 159 |
+
update = update.mT
|
| 160 |
+
update = update / (update.norm(dim=(-2, -1), keepdim=True) * (1 + 2e-2) + 1e-6)
|
| 161 |
+
for a, b, c in NS_COEFFS:
|
| 162 |
+
gram = update @ update.mT
|
| 163 |
+
polynomial = b * gram + c * gram @ gram
|
| 164 |
+
update = a * update + polynomial @ update
|
| 165 |
+
if transpose:
|
| 166 |
+
update = update.mT
|
| 167 |
+
p.mul_(1 - lr * MUON_WD)
|
| 168 |
+
# lr may be a 0-d tensor (CUDA-graphed step); scale in p's dtype like add_(alpha=-lr)
|
| 169 |
+
p.sub_(lr * update.reshape_as(p).to(p.dtype))
|
| 170 |
+
|
| 171 |
+
|
| 172 |
+
class MuonStep:
|
| 173 |
+
"""muon_step on `matrices` with fresh momentum buffers, called once per step after backward
|
| 174 |
+
(grads must keep fixed addresses: zero_grad(set_to_none=False)). On CUDA the first call
|
| 175 |
+
captures the step as one CUDA graph and later calls replay it instead of ~400 kernel
|
| 176 |
+
launches; CPU smoke runs call muon_step directly."""
|
| 177 |
+
|
| 178 |
+
def __init__(self, matrices, device):
|
| 179 |
+
self.matrices, self.device = matrices, device
|
| 180 |
+
self.buffers = [torch.zeros_like(p) for p in matrices]
|
| 181 |
+
self.lr, self.graph = torch.zeros((), device=device), None
|
| 182 |
+
|
| 183 |
+
def __call__(self, lr):
|
| 184 |
+
self.lr.fill_(lr)
|
| 185 |
+
if self.device != "cuda":
|
| 186 |
+
muon_step(self.matrices, self.buffers, self.lr)
|
| 187 |
+
return
|
| 188 |
+
if self.graph is None:
|
| 189 |
+
state = self.matrices + self.buffers
|
| 190 |
+
saved = [t.clone() for t in state]
|
| 191 |
+
side = torch.cuda.Stream()
|
| 192 |
+
side.wait_stream(torch.cuda.current_stream())
|
| 193 |
+
with torch.cuda.stream(side):
|
| 194 |
+
muon_step(self.matrices, self.buffers, self.lr) # warm-up (lazy cuBLAS init)
|
| 195 |
+
torch.cuda.current_stream().wait_stream(side)
|
| 196 |
+
self.graph = torch.cuda.CUDAGraph()
|
| 197 |
+
with torch.cuda.graph(self.graph):
|
| 198 |
+
muon_step(self.matrices, self.buffers, self.lr)
|
| 199 |
+
with torch.no_grad():
|
| 200 |
+
for t, v in zip(state, saved, strict=True):
|
| 201 |
+
t.copy_(v) # undo the warm-up and capture updates
|
| 202 |
+
self.graph.replay()
|
| 203 |
+
|
| 204 |
+
|
| 205 |
+
class PeriLNBlock(Block):
|
| 206 |
+
"""P22 Peri-LN (Kim et al. 2025): the fork's pre-LN block plus a LayerNorm on each branch
|
| 207 |
+
output, so every residual write has unit scale times a learned gain. Attention has no RoPE:
|
| 208 |
+
`mask` is the additive ALiBi bias (B or 1, H, T, S) that LoopedGPT builds once per forward
|
| 209 |
+
and cos/sin are unused."""
|
| 210 |
+
|
| 211 |
+
def __init__(self, d, h):
|
| 212 |
+
super().__init__(d, h)
|
| 213 |
+
self.fc1, self.fc2 = torch.nn.Linear(d, D_FF), torch.nn.Linear(D_FF, d)
|
| 214 |
+
self.ln_attn_out, self.ln_mlp_out = torch.nn.LayerNorm(d), torch.nn.LayerNorm(d)
|
| 215 |
+
|
| 216 |
+
def attn(self, x, cos, sin, mask, cache):
|
| 217 |
+
B, T, D = x.shape
|
| 218 |
+
q, k, v = self.qkv(x).view(B, T, 3, self.h, D // self.h).permute(2, 0, 3, 1, 4)
|
| 219 |
+
if cache is not None: # the fork's preallocated KV cache, unchanged
|
| 220 |
+
if "k" not in cache:
|
| 221 |
+
cache["k"] = k.new_empty(B, self.h, cache["max"], D // self.h)
|
| 222 |
+
cache["v"] = v.new_empty(B, self.h, cache["max"], D // self.h)
|
| 223 |
+
cache["n"] = 0
|
| 224 |
+
n0, n1 = cache["n"], cache["n"] + T
|
| 225 |
+
cache["k"][:, :, n0:n1] = k
|
| 226 |
+
cache["v"][:, :, n0:n1] = v
|
| 227 |
+
cache["n"] = n1
|
| 228 |
+
k, v = cache["k"][:, :, :n1], cache["v"][:, :, :n1]
|
| 229 |
+
y = torch.nn.functional.scaled_dot_product_attention(q, k, v, attn_mask=mask.to(q.dtype))
|
| 230 |
+
return self.proj(y.transpose(1, 2).reshape(B, T, D))
|
| 231 |
+
|
| 232 |
+
def forward(self, x, cos, sin, mask=None, cache=None):
|
| 233 |
+
x = x + self.ln_attn_out(self.attn(self.ln1(x), cos, sin, mask, cache))
|
| 234 |
+
return x + self.ln_mlp_out(self.fc2(torch.nn.functional.gelu(self.fc1(self.ln2(x)))))
|
| 235 |
+
|
| 236 |
+
|
| 237 |
+
class LoopedGPT(GPT):
|
| 238 |
+
"""Looped trunk (Universal Transformer, Dehghani et al. 2018; sandwich layout and input
|
| 239 |
+
injection as in Geiping et al. 2025): blocks[0] is the entry block, blocks[1:-1] the shared
|
| 240 |
+
core applied n times, blocks[-1] the exit block. Every core pass starts from
|
| 241 |
+
ln_loop(state) + e, with e the entry block's output, so each pass sees an input of the same
|
| 242 |
+
scale and the prompt is re-injected instead of being buried under accumulated writes.
|
| 243 |
+
|
| 244 |
+
P34: the fork's GPT with ALiBi (Press et al. 2021) instead of RoPE. Head h's attention
|
| 245 |
+
logits get -slope_h * (query position - key position), slopes 2^-1 .. 2^-8 for 8 heads, so
|
| 246 |
+
no rotation of q/k anywhere. Positions come from `pos` (sample.generate's left-padded
|
| 247 |
+
batches) or arange; key positions are kept in caches[0] so cached decoding sees the same
|
| 248 |
+
bias as the full forward. Causality/padding come from `mask` when given, else from rel < 0."""
|
| 249 |
+
|
| 250 |
+
def __init__(self, vocab, n_layer, d, n_head, seed, max_len=1024):
|
| 251 |
+
super().__init__(vocab, n_layer, d, n_head, max_len)
|
| 252 |
+
slopes = torch.tensor(ALIBI_SLOPES).view(1, n_head, 1, 1)
|
| 253 |
+
self.register_buffer("slopes", slopes, persistent=False)
|
| 254 |
+
self.ln_loop = torch.nn.LayerNorm(d)
|
| 255 |
+
self.n_loops, self.loop_rng = DEFAULT_LOOPS, random.Random(seed)
|
| 256 |
+
|
| 257 |
+
def forward(self, idx, pos=None, mask=None, caches=None):
|
| 258 |
+
return self.head(self.ln_f(self.features(idx, pos, mask, caches)[0]))
|
| 259 |
+
|
| 260 |
+
def features(self, idx, pos=None, mask=None, caches=None):
|
| 261 |
+
"""The trunk: (final hidden state before ln_f, the ALiBi bias used)."""
|
| 262 |
+
B, T = idx.shape
|
| 263 |
+
if pos is None: # training/val: one (1, H, T, T) bias broadcast over the batch
|
| 264 |
+
pos = torch.arange(T, device=idx.device)[None]
|
| 265 |
+
key_pos = pos
|
| 266 |
+
if caches is not None: # incremental decoding: keys are every cached token
|
| 267 |
+
c = caches[0]
|
| 268 |
+
if "pos" not in c:
|
| 269 |
+
c["pos"], c["n_pos"] = pos.new_empty(pos.size(0), c["max"]), 0
|
| 270 |
+
n0, n1 = c["n_pos"], c["n_pos"] + T
|
| 271 |
+
c["pos"][:, n0:n1] = pos
|
| 272 |
+
c["n_pos"] = n1
|
| 273 |
+
key_pos = c["pos"][:, :n1]
|
| 274 |
+
rel = pos[:, None, :, None] - key_pos[:, None, None, :] # (B or 1, 1, T, S)
|
| 275 |
+
bias = -self.slopes * rel.to(self.slopes.dtype)
|
| 276 |
+
bias = bias.masked_fill(rel < 0 if mask is None else ~mask, float("-inf"))
|
| 277 |
+
n = self.loop_rng.choice(TRAIN_LOOPS) if self.training else self.n_loops
|
| 278 |
+
entry, *core, exit_block = self.blocks
|
| 279 |
+
x = e = entry(self.emb(idx), None, None, bias, None if caches is None else caches[0])
|
| 280 |
+
for loop in range(n):
|
| 281 |
+
x = self.ln_loop(x) + e
|
| 282 |
+
for j, b in enumerate(core, 1):
|
| 283 |
+
# sample.generate makes one cache dict per unique block; a shared block's keys
|
| 284 |
+
# and values differ per pass, so each pass gets its own sub-cache inside it.
|
| 285 |
+
c = caches and caches[j].setdefault(loop, {"max": caches[j]["max"]})
|
| 286 |
+
x = b(x, None, None, bias, c)
|
| 287 |
+
x = exit_block(x, None, None, bias, None if caches is None else caches[-1])
|
| 288 |
+
return x, bias
|
| 289 |
+
|
| 290 |
+
|
| 291 |
+
class MTPHead(torch.nn.Module):
|
| 292 |
+
"""The MTP module. A separate nn.Module, never registered on the model, so the harness's
|
| 293 |
+
parameter count, RL AdamW and sampler never see it; it is dropped when pretrain() returns."""
|
| 294 |
+
|
| 295 |
+
def __init__(self, d, h):
|
| 296 |
+
super().__init__()
|
| 297 |
+
self.ln_h, self.ln_e = torch.nn.LayerNorm(d), torch.nn.LayerNorm(d)
|
| 298 |
+
self.proj = torch.nn.Linear(2 * d, d, bias=False)
|
| 299 |
+
self.block = PeriLNBlock(d, h)
|
| 300 |
+
|
| 301 |
+
|
| 302 |
+
def mtp_loss(model, mtp, x, m):
|
| 303 |
+
"""The fork's loss_on (next-token CE on proof tokens) plus MTP_WEIGHT times the CE of the
|
| 304 |
+
MTP module's token-(t+2) prediction at the same proof positions. Returns (total, main)."""
|
| 305 |
+
h, bias = model.features(x[:, :-1])
|
| 306 |
+
logits = model.head(model.ln_f(h))
|
| 307 |
+
V = logits.size(-1)
|
| 308 |
+
ce = torch.nn.functional.cross_entropy(
|
| 309 |
+
logits.reshape(-1, V).float(), x[:, 1:].reshape(-1), reduction="none"
|
| 310 |
+
)
|
| 311 |
+
main = (ce * m[:, 1:].reshape(-1)).sum() / m[:, 1:].sum()
|
| 312 |
+
# Position t (0..T-3) gets h_t and emb(x_{t+1}); causal attention over these keeps every
|
| 313 |
+
# prediction of x_{t+2} a function of x_{<=t+1}.
|
| 314 |
+
z = mtp.proj(torch.cat((mtp.ln_h(h[:, :-1]), mtp.ln_e(model.emb(x[:, 1:-1]))), -1))
|
| 315 |
+
z = mtp.block(z, None, None, bias[..., :-1, :-1], None)
|
| 316 |
+
ce2 = torch.nn.functional.cross_entropy(
|
| 317 |
+
model.head(model.ln_f(z)).reshape(-1, V).float(), x[:, 2:].reshape(-1), reduction="none"
|
| 318 |
+
)
|
| 319 |
+
aux = (ce2 * m[:, 2:].reshape(-1)).sum() / m[:, 2:].sum().clamp(min=1)
|
| 320 |
+
return main + MTP_WEIGHT * aux, main, aux
|
| 321 |
+
|
| 322 |
+
|
| 323 |
+
def val_loss(model, ids, mask, ref_max, lens, tok, gen, autocast):
|
| 324 |
+
"""Mean loss over the first VAL_N records (never trained on), no shift augmentation."""
|
| 325 |
+
model.eval()
|
| 326 |
+
tot = 0.0
|
| 327 |
+
with torch.no_grad(), autocast:
|
| 328 |
+
for i in range(0, VAL_N, BS):
|
| 329 |
+
idxs = torch.arange(i, min(i + BS, VAL_N), device=ids.device)
|
| 330 |
+
x, m = ids[idxs], mask[idxs]
|
| 331 |
+
T = int(lens[idxs].max())
|
| 332 |
+
tot += loss_on(model, x[:, :T], m[:, :T]).item() * len(idxs)
|
| 333 |
+
model.train()
|
| 334 |
+
return tot / VAL_N
|
| 335 |
+
|
| 336 |
+
|
| 337 |
+
def to_gpu(data, tok, device):
|
| 338 |
+
"""Pad all (prompt, proof) records once into (N, T) ids + proof mask on the device."""
|
| 339 |
+
N, T = len(data), max(len(p) + len(q) for p, q in data)
|
| 340 |
+
ids = torch.full((N, T), tok.pad, dtype=torch.long)
|
| 341 |
+
mask = torch.zeros((N, T), dtype=torch.bool)
|
| 342 |
+
ref_max = torch.zeros(
|
| 343 |
+
N, dtype=torch.long
|
| 344 |
+
) # largest N<k> index used in the proof, for the shift augmentation
|
| 345 |
+
for i, (p, q) in enumerate(data):
|
| 346 |
+
ids[i, : len(p) + len(q)] = torch.tensor(p + q)
|
| 347 |
+
mask[i, len(p) : len(p) + len(q)] = True
|
| 348 |
+
ref_max[i] = max((x - tok.ref0 + 1 for x in q if x >= tok.ref0), default=0)
|
| 349 |
+
return ids.to(device), mask.to(device), ref_max.to(device), (ids != tok.pad).sum(1).to(device)
|
| 350 |
+
|
| 351 |
+
|
| 352 |
+
def token_batches(lens, n_train, gen, rng, first=VAL_N):
|
| 353 |
+
"""One epoch of token-budget batches: a random permutation of the n_train records from
|
| 354 |
+
`first` on, stably sorted by length (equal lengths stay shuffled), cut greedily into runs of
|
| 355 |
+
consecutive records with count * max_len <= TOK_BUDGET, and the runs' order shuffled.
|
| 356 |
+
Returns (order, cuts)."""
|
| 357 |
+
perm = torch.randperm(n_train, device=lens.device, generator=gen) + first
|
| 358 |
+
order = perm[torch.argsort(lens[perm], stable=True)]
|
| 359 |
+
sorted_lens = lens[order].cpu()
|
| 360 |
+
cuts, s = [], 0
|
| 361 |
+
while s < len(sorted_lens):
|
| 362 |
+
window = sorted_lens[s : s + TOK_BUDGET // int(sorted_lens[s])]
|
| 363 |
+
# count * max_len is non-decreasing along the window, so the fits form a prefix.
|
| 364 |
+
e = s + int((torch.arange(1, len(window) + 1) * window <= TOK_BUDGET).sum())
|
| 365 |
+
cuts.append((s, e, int(sorted_lens[e - 1]))) # sorted: last is the batch's max length
|
| 366 |
+
s = e
|
| 367 |
+
rng.shuffle(cuts)
|
| 368 |
+
return order, cuts
|
| 369 |
+
|
| 370 |
+
|
| 371 |
+
def gpu_batch(ids, mask, ref_max, lens, idxs, tok, gen, max_len=None):
|
| 372 |
+
"""The fork's batch() + tok.shift_abs, vectorised: a random per-row offset is added to
|
| 373 |
+
every ref token."""
|
| 374 |
+
x, m = ids[idxs], mask[idxs]
|
| 375 |
+
# Known on the host for training batches: passing it avoids a GPU sync every step.
|
| 376 |
+
T = int(lens[idxs].max()) if max_len is None else max_len
|
| 377 |
+
x, m = x[:, :T].clone(), m[:, :T]
|
| 378 |
+
if tok.mode in ("abs", "lean_seq"): # lean_seq: names n1.. get the same start offset
|
| 379 |
+
room = MAXN - ref_max[idxs]
|
| 380 |
+
shift = (
|
| 381 |
+
(torch.rand(len(idxs), device=x.device, generator=gen) * (room + 1))
|
| 382 |
+
.long()
|
| 383 |
+
.clamp(max=room)
|
| 384 |
+
)
|
| 385 |
+
is_ref = (x >= tok.ref0) & m
|
| 386 |
+
x = torch.where(is_ref, x + shift[:, None], x)
|
| 387 |
+
return x, m
|
| 388 |
+
|
| 389 |
+
|
| 390 |
+
def pretrain(train_recs, budget_secs, device, seed, log):
|
| 391 |
+
t0 = time.time()
|
| 392 |
+
tok = make_tokenizer(TOK_MODE)
|
| 393 |
+
data = [(tok.encode_prompt(r["prompt"]), tok.encode_proof(r["proof"])) for r in train_recs]
|
| 394 |
+
ids, mask, ref_max, lens = to_gpu(data, tok, device)
|
| 395 |
+
gen = torch.Generator(device=device)
|
| 396 |
+
gen.manual_seed(seed)
|
| 397 |
+
model = LoopedGPT(tok.vocab_size, N_LAYER, D, N_HEAD, seed)
|
| 398 |
+
model.blocks = torch.nn.ModuleList(PeriLNBlock(D, N_HEAD) for _ in range(N_LAYER))
|
| 399 |
+
model.blocks.apply(model._init) # fork init: zero biases
|
| 400 |
+
model = model.to(device)
|
| 401 |
+
mtp = MTPHead(D, N_HEAD)
|
| 402 |
+
mtp.apply(model._init)
|
| 403 |
+
mtp = mtp.to(device)
|
| 404 |
+
# Compile the repeated blocks once each (torch >= 2.9; dynamic shapes, ~14 s): ~14% faster
|
| 405 |
+
# steps. The coalescing tiling analysis crashes on dynamic shapes, so it is off. Undone at
|
| 406 |
+
# handoff so the harness's RL fine-tune and KV-cache sampling run eager.
|
| 407 |
+
torch._inductor.config.triton.coalesce_tiling_analysis = False
|
| 408 |
+
compiled = list(model.blocks) + [mtp.block]
|
| 409 |
+
for b in compiled if device == "cuda" else []: # CPU smoke runs stay eager
|
| 410 |
+
b.compile(dynamic=True)
|
| 411 |
+
for p in list(model.blocks.parameters()) + list(mtp.parameters()):
|
| 412 |
+
if p.ndim == 2:
|
| 413 |
+
# Orthogonal matrices have element second moment gain^2 / max(rows, cols).
|
| 414 |
+
torch.nn.init.orthogonal_(p, gain=0.02 * math.sqrt(max(p.shape)))
|
| 415 |
+
log(
|
| 416 |
+
f"pretrain: {model.n_params():,} params, {len(data):,} records on {device} "
|
| 417 |
+
f"after {time.time() - t0:.0f}s, budget {budget_secs}s"
|
| 418 |
+
)
|
| 419 |
+
# The output readout joins the hidden matrices under Muon; embeddings stay on AdamW.
|
| 420 |
+
trainable = list(model.parameters()) + list(mtp.parameters())
|
| 421 |
+
matrices = [p for p in model.blocks.parameters() if p.ndim == 2] + [model.head.weight]
|
| 422 |
+
matrices += [p for p in mtp.parameters() if p.ndim == 2]
|
| 423 |
+
matrix_ids = {id(p) for p in matrices}
|
| 424 |
+
auxiliary = [p for p in trainable if id(p) not in matrix_ids]
|
| 425 |
+
muon = MuonStep(matrices, device)
|
| 426 |
+
opt = torch.optim.AdamW(
|
| 427 |
+
auxiliary, lr=LR, weight_decay=WD, betas=(0.9, 0.95), fused=device == "cuda"
|
| 428 |
+
)
|
| 429 |
+
autocast = torch.autocast("cuda", dtype=torch.bfloat16, enabled=(device == "cuda"))
|
| 430 |
+
model.train()
|
| 431 |
+
step, last_loss, records_seen = 0, float("nan"), 0
|
| 432 |
+
curve, recent, recent_aux, next_point = [], [], [], CURVE_EVERY
|
| 433 |
+
n_train = len(data) - VAL_N # records [0, VAL_N) are the val slice
|
| 434 |
+
rng, cuts, batches_per_epoch = random.Random(seed), [], None
|
| 435 |
+
order = torch.empty(0, dtype=torch.long, device=device)
|
| 436 |
+
while True:
|
| 437 |
+
frac = (
|
| 438 |
+
time.time() - t0
|
| 439 |
+
) / budget_secs # cosine over wall-clock, so the cap is the schedule
|
| 440 |
+
if frac >= 1.0:
|
| 441 |
+
break
|
| 442 |
+
step += 1
|
| 443 |
+
# No warmup (as in modded-nanogpt): Muon's step is normalized and Adam's is ~lr per
|
| 444 |
+
# element from step 1, and Peri-LN normalizes every branch write.
|
| 445 |
+
lr = LR * MIN_LR_FRAC + 0.5 * LR * (1 - MIN_LR_FRAC) * (1 + math.cos(math.pi * frac))
|
| 446 |
+
for g in opt.param_groups:
|
| 447 |
+
g["lr"] = lr
|
| 448 |
+
if not cuts: # new epoch
|
| 449 |
+
order, cuts = token_batches(lens, n_train, gen, rng)
|
| 450 |
+
batches_per_epoch = len(cuts)
|
| 451 |
+
s, e, T = cuts.pop()
|
| 452 |
+
idxs = order[s:e]
|
| 453 |
+
records_seen += e - s
|
| 454 |
+
x, m = gpu_batch(ids, mask, ref_max, lens, idxs, tok, gen, T)
|
| 455 |
+
with autocast:
|
| 456 |
+
loss, main, aux = mtp_loss(model, mtp, x, m)
|
| 457 |
+
model.zero_grad(set_to_none=False) # grads keep fixed addresses for the Muon CUDA graph
|
| 458 |
+
mtp.zero_grad(set_to_none=False)
|
| 459 |
+
loss.backward()
|
| 460 |
+
torch.nn.utils.clip_grad_norm_(trainable, 1.0)
|
| 461 |
+
opt.step()
|
| 462 |
+
muon(MUON_LR * lr / LR)
|
| 463 |
+
if step % 20 == 0:
|
| 464 |
+
last_loss = main.item() # the next-token CE, comparable with earlier runs
|
| 465 |
+
recent.append(last_loss)
|
| 466 |
+
recent_aux.append(aux.item())
|
| 467 |
+
if (
|
| 468 |
+
time.time() - t0 >= next_point
|
| 469 |
+
): # curve point: mean recent train loss + val loss on the held-out slice
|
| 470 |
+
curve.append(
|
| 471 |
+
{
|
| 472 |
+
"secs": round(time.time() - t0),
|
| 473 |
+
"step": step,
|
| 474 |
+
"train_loss": sum(recent) / len(recent) if recent else last_loss,
|
| 475 |
+
"val_loss": val_loss(model, ids, mask, ref_max, lens, tok, gen, autocast),
|
| 476 |
+
"mtp_loss": sum(recent_aux) / len(recent_aux) if recent_aux else None,
|
| 477 |
+
}
|
| 478 |
+
)
|
| 479 |
+
log(
|
| 480 |
+
f"pretrain step {step} train {curve[-1]['train_loss']:.4f} "
|
| 481 |
+
f"val {curve[-1]['val_loss']:.4f} lr {lr:.2e} {curve[-1]['secs']}s"
|
| 482 |
+
)
|
| 483 |
+
recent, recent_aux, next_point = [], [], next_point + CURVE_EVERY
|
| 484 |
+
val_by_loops = {}
|
| 485 |
+
for n in VAL_LOOPS if step else (): # no steps: a checkpoint rebuild, not a training run
|
| 486 |
+
model.n_loops = n
|
| 487 |
+
val_by_loops[n] = val_loss(model, ids, mask, ref_max, lens, tok, gen, autocast)
|
| 488 |
+
model.n_loops = DEFAULT_LOOPS
|
| 489 |
+
log(f"pretrain val loss by loops: {val_by_loops}")
|
| 490 |
+
for b in compiled:
|
| 491 |
+
b._compiled_call_impl = None
|
| 492 |
+
model.eval()
|
| 493 |
+
return (
|
| 494 |
+
model,
|
| 495 |
+
tok,
|
| 496 |
+
{
|
| 497 |
+
"steps": step,
|
| 498 |
+
"records_seen": records_seen,
|
| 499 |
+
"epochs": records_seen / n_train,
|
| 500 |
+
"batches_per_epoch": batches_per_epoch,
|
| 501 |
+
"tok_budget": TOK_BUDGET,
|
| 502 |
+
"last_loss": last_loss,
|
| 503 |
+
"train_loops": TRAIN_LOOPS,
|
| 504 |
+
"default_loops": DEFAULT_LOOPS,
|
| 505 |
+
"val_by_loops": val_by_loops,
|
| 506 |
+
"curve": curve,
|
| 507 |
+
},
|
| 508 |
+
)
|
| 509 |
+
|
| 510 |
+
|
| 511 |
+
def finetune(model, tok, data, rng, device):
|
| 512 |
+
"""The RL fine-tune: pretraining's optimizers (Muon on the hidden matrices, AdamW on the
|
| 513 |
+
rest), linear warmup then cosine to lr/10, over FT_RECORDS records in token-budget batches
|
| 514 |
+
(to_gpu / token_batches / gpu_batch, as pretraining). Muon's momentum restarts every round,
|
| 515 |
+
as AdamW's moments did. The model is in train mode, so every batch draws its own loop count
|
| 516 |
+
from TRAIN_LOOPS. Returns (last loss, steps)."""
|
| 517 |
+
ids, mask, ref_max, lens = to_gpu(data, tok, device)
|
| 518 |
+
gen = torch.Generator(device=device)
|
| 519 |
+
gen.manual_seed(rng.getrandbits(31))
|
| 520 |
+
plan, seen = [], 0 # (order, start, end, padded length) per step, epoch after epoch
|
| 521 |
+
while seen < FT_RECORDS:
|
| 522 |
+
order, cuts = token_batches(lens, len(data), gen, rng, first=0)
|
| 523 |
+
for s, e, T in cuts:
|
| 524 |
+
if seen >= FT_RECORDS:
|
| 525 |
+
break
|
| 526 |
+
plan.append((order, s, e, T))
|
| 527 |
+
seen += e - s
|
| 528 |
+
matrices = [p for p in model.blocks.parameters() if p.ndim == 2] + [model.head.weight]
|
| 529 |
+
matrix_ids = {id(p) for p in matrices}
|
| 530 |
+
auxiliary = [p for p in model.parameters() if id(p) not in matrix_ids]
|
| 531 |
+
opt = torch.optim.AdamW(auxiliary, lr=FT_LR, weight_decay=0.1, betas=(0.9, 0.95))
|
| 532 |
+
muon = MuonStep(matrices, device)
|
| 533 |
+
autocast = torch.autocast("cuda", dtype=torch.bfloat16, enabled=(device == "cuda"))
|
| 534 |
+
model.train()
|
| 535 |
+
for step, (order, s, e, T) in enumerate(plan, 1):
|
| 536 |
+
lr = (
|
| 537 |
+
FT_LR * step / FT_WARMUP
|
| 538 |
+
if step < FT_WARMUP
|
| 539 |
+
else FT_LR / 10 + 0.45 * FT_LR * (1 + math.cos(math.pi * step / len(plan)))
|
| 540 |
+
)
|
| 541 |
+
for g in opt.param_groups:
|
| 542 |
+
g["lr"] = lr
|
| 543 |
+
x, m = gpu_batch(ids, mask, ref_max, lens, order[s:e], tok, gen, T)
|
| 544 |
+
with autocast:
|
| 545 |
+
loss = loss_on(model, x, m)
|
| 546 |
+
model.zero_grad(set_to_none=False) # grads keep fixed addresses for the Muon CUDA graph
|
| 547 |
+
loss.backward()
|
| 548 |
+
torch.nn.utils.clip_grad_norm_(model.parameters(), 1.0)
|
| 549 |
+
opt.step()
|
| 550 |
+
muon(FT_MUON_LR * lr / FT_LR)
|
| 551 |
+
model.eval()
|
| 552 |
+
return loss.item(), len(plan)
|
| 553 |
+
|
| 554 |
+
|
| 555 |
+
def rl(model, tok, targets, train_recs, oracle, budget_secs, device, seed, log):
|
| 556 |
+
rng = random.Random(seed)
|
| 557 |
+
found = {t["name"]: {} for t in targets} # name -> {normalized proof: proof text}
|
| 558 |
+
rounds, round_secs, retries, ft_steps = 0, 0.0, [], []
|
| 559 |
+
# Start a round only if one as long as the last (plus 15%) still fits in the budget.
|
| 560 |
+
while oracle.left() > 1.15 * round_secs:
|
| 561 |
+
t0 = oracle.used()
|
| 562 |
+
model.n_loops = RL_FIRST_LOOPS
|
| 563 |
+
rows = oracle.sample(model, targets, RL_K)
|
| 564 |
+
failed = [t for t, row in zip(targets, rows, strict=True) if not row["proofs"]]
|
| 565 |
+
rng.shuffle(failed)
|
| 566 |
+
for i, n in enumerate(RL_RETRY_LOOPS):
|
| 567 |
+
part = failed[i :: len(RL_RETRY_LOOPS)]
|
| 568 |
+
if not part:
|
| 569 |
+
continue
|
| 570 |
+
model.n_loops = n
|
| 571 |
+
part_rows = oracle.sample(model, part, RL_RETRY_K)
|
| 572 |
+
cracked = sum(bool(row["proofs"]) for row in part_rows)
|
| 573 |
+
retries.append(
|
| 574 |
+
{"round": rounds + 1, "loops": n, "targets": len(part), "cracked": cracked}
|
| 575 |
+
)
|
| 576 |
+
rows += part_rows
|
| 577 |
+
for row in rows:
|
| 578 |
+
for proof in row["proofs"]:
|
| 579 |
+
found[row["name"]].setdefault(norm(proof), proof)
|
| 580 |
+
mix = []
|
| 581 |
+
for t in targets:
|
| 582 |
+
proofs = list(found[t["name"]].values())
|
| 583 |
+
rng.shuffle(proofs)
|
| 584 |
+
mix += [(t["prompt"], proof) for proof in proofs[:MAX_PER_THM]] * RL_WEIGHT
|
| 585 |
+
n_rl = len(mix)
|
| 586 |
+
if not n_rl:
|
| 587 |
+
break
|
| 588 |
+
mix += [(r["prompt"], r["proof"]) for r in rng.sample(train_recs, RETAIN)]
|
| 589 |
+
data = [(tok.encode_prompt(p), tok.encode_proof(q)) for p, q in mix]
|
| 590 |
+
loss, steps = finetune(model, tok, data, rng, device)
|
| 591 |
+
ft_steps.append(steps)
|
| 592 |
+
rounds += 1
|
| 593 |
+
round_secs = oracle.used() - t0
|
| 594 |
+
log(
|
| 595 |
+
f"rl round {rounds}: {n_rl} RL records, fine-tune loss {loss:.4f} in {steps} steps, "
|
| 596 |
+
f"round {round_secs:.0f}s, {oracle.left():.0f}s left"
|
| 597 |
+
)
|
| 598 |
+
model.n_loops = DEFAULT_LOOPS
|
| 599 |
+
return {
|
| 600 |
+
"rounds": rounds,
|
| 601 |
+
"first_loops": RL_FIRST_LOOPS,
|
| 602 |
+
"k_first": RL_K,
|
| 603 |
+
"retry_loops": RL_RETRY_LOOPS,
|
| 604 |
+
"k_retry": RL_RETRY_K,
|
| 605 |
+
"retries": retries,
|
| 606 |
+
"ft_records": FT_RECORDS,
|
| 607 |
+
"ft_steps": ft_steps,
|
| 608 |
+
"ft_optimizer": "muon (polar express) on block matrices and head + adamw on the rest",
|
| 609 |
+
"ft_lr": FT_LR,
|
| 610 |
+
"ft_muon_lr": FT_MUON_LR,
|
| 611 |
+
}
|
runs/510-fable_more-shorter-rounds_s1/source/remote.sh
ADDED
|
@@ -0,0 +1,29 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
#!/usr/bin/env bash
|
| 2 |
+
# Runs on the pod. Checks out Dan's fork at the pinned commit (frozen data + verifier + Lean
|
| 3 |
+
# tokenizer/gate), installs Lean, runs the harness.
|
| 4 |
+
# Usage: bash remote.sh <out-dir> [harness.py args...]
|
| 5 |
+
set -euo pipefail
|
| 6 |
+
FORK_COMMIT=51604b3feaa5e66821d1faba22238fe6d96d46d0 # dan-pandori/nd-takehome, branch dan_lean_seed2 (lean_seq with the BOTE fix)
|
| 7 |
+
export ND_FORK=/workspace/nd-takehome
|
| 8 |
+
OUT=$1; shift
|
| 9 |
+
|
| 10 |
+
if [ ! -d "$ND_FORK/.git" ]; then
|
| 11 |
+
git clone -q https://github.com/dan-pandori/nd-takehome.git "$ND_FORK"
|
| 12 |
+
fi
|
| 13 |
+
git -C "$ND_FORK" cat-file -e "$FORK_COMMIT" 2>/dev/null || git -C "$ND_FORK" fetch -q origin
|
| 14 |
+
git -C "$ND_FORK" checkout -q "$FORK_COMMIT"
|
| 15 |
+
[ -f "$ND_FORK/data/train.jsonl" ] || gunzip -k "$ND_FORK/data/train.jsonl.gz"
|
| 16 |
+
# Lean 4 core checks every sampled proof (fork lean_gate.py, called from sample.generate): the
|
| 17 |
+
# reward is Lean AND nd_verify. Installed before harness.py starts, so it is not charged to TOTAL_SECS.
|
| 18 |
+
[ -x ~/.elan/bin/lean ] || curl -sSf https://raw.githubusercontent.com/leanprover/elan/master/elan-init.sh \
|
| 19 |
+
| sh -s -- -y --default-toolchain leanprover/lean4:v4.34.0 > /dev/null
|
| 20 |
+
# One Lean process per CPU: on these pods (nproc 96) 12 workers took 146 s per RL-round gate
|
| 21 |
+
# pass (099) against 67-84 s with 96 workers (098).
|
| 22 |
+
export LEAN_GATE_WORKERS=$(nproc) LEAN_GATE_LOG="$OUT/lean_gate.jsonl"
|
| 23 |
+
nvidia-smi --query-gpu=name,memory.total --format=csv,noheader
|
| 24 |
+
echo "cpus $(nproc), $(~/.elan/bin/lean --version)"
|
| 25 |
+
# Newer RunPod images put torch under `python` and leave `python3` as bare system Python.
|
| 26 |
+
PY=$(for p in python python3; do "$p" -c "import torch" 2>/dev/null && echo "$p" && break; done)
|
| 27 |
+
echo "python $($PY --version 2>&1), torch $($PY -c 'import torch; print(torch.__version__)')"
|
| 28 |
+
cd /workspace/autoresearch
|
| 29 |
+
$PY harness.py --out "$OUT" "$@" 2>&1 | tee "$OUT.log"
|
runs/510-fable_more-shorter-rounds_s1/summary.json
ADDED
|
@@ -0,0 +1,2321 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"status": "ok",
|
| 3 |
+
"metric": 815,
|
| 4 |
+
"metric_diag": 795,
|
| 5 |
+
"metric_rl": 1314,
|
| 6 |
+
"harness_sha": "2eec1b61666a",
|
| 7 |
+
"pretrain_sha": "6c50c444778d",
|
| 8 |
+
"device": "cuda",
|
| 9 |
+
"gpu": "NVIDIA A40",
|
| 10 |
+
"torch": "2.9.1+cu129",
|
| 11 |
+
"lean_gate": {
|
| 12 |
+
"samples": 445656,
|
| 13 |
+
"parse_fail": 99756,
|
| 14 |
+
"distinct_checked": 323460,
|
| 15 |
+
"both_ok": 165135,
|
| 16 |
+
"nd_ok_lean_rej": 0,
|
| 17 |
+
"nd_rej_lean_ok": 19,
|
| 18 |
+
"lean_wall_s": 956.8983767032623
|
| 19 |
+
},
|
| 20 |
+
"total_secs": 3941.4172530174255,
|
| 21 |
+
"rules": {
|
| 22 |
+
"TOTAL_SECS": 5400,
|
| 23 |
+
"PRETRAIN_SECS": 300,
|
| 24 |
+
"RL_SECS": 1800,
|
| 25 |
+
"PARAM_CAP": 10000000,
|
| 26 |
+
"N_TARGETS": 1500,
|
| 27 |
+
"EVAL_K": 64,
|
| 28 |
+
"KNOB": "n_loops",
|
| 29 |
+
"EVAL_LEVEL": 8,
|
| 30 |
+
"DIAG_LEVEL": 2,
|
| 31 |
+
"TOK_MODE": "lean_seq"
|
| 32 |
+
},
|
| 33 |
+
"pretrain": {
|
| 34 |
+
"steps": 3470,
|
| 35 |
+
"records_seen": 471748,
|
| 36 |
+
"epochs": 3.063975163347752,
|
| 37 |
+
"batches_per_epoch": 1132,
|
| 38 |
+
"tok_budget": 18432,
|
| 39 |
+
"last_loss": 0.05730755999684334,
|
| 40 |
+
"train_loops": [
|
| 41 |
+
1,
|
| 42 |
+
2,
|
| 43 |
+
3,
|
| 44 |
+
4,
|
| 45 |
+
5,
|
| 46 |
+
6,
|
| 47 |
+
7,
|
| 48 |
+
8
|
| 49 |
+
],
|
| 50 |
+
"default_loops": 2,
|
| 51 |
+
"val_by_loops": {
|
| 52 |
+
"1": 0.09079534560441971,
|
| 53 |
+
"2": 0.08942204620689154,
|
| 54 |
+
"4": 0.08907637745141983,
|
| 55 |
+
"8": 0.08901151362806559
|
| 56 |
+
},
|
| 57 |
+
"curve": [
|
| 58 |
+
{
|
| 59 |
+
"secs": 30,
|
| 60 |
+
"step": 1,
|
| 61 |
+
"train_loss": NaN,
|
| 62 |
+
"val_loss": 3.9689463675022125,
|
| 63 |
+
"mtp_loss": null
|
| 64 |
+
},
|
| 65 |
+
{
|
| 66 |
+
"secs": 32,
|
| 67 |
+
"step": 2,
|
| 68 |
+
"train_loss": NaN,
|
| 69 |
+
"val_loss": 3.4865831434726715,
|
| 70 |
+
"mtp_loss": null
|
| 71 |
+
},
|
| 72 |
+
{
|
| 73 |
+
"secs": 45,
|
| 74 |
+
"step": 170,
|
| 75 |
+
"train_loss": 0.560400852933526,
|
| 76 |
+
"val_loss": 0.13067163340747356,
|
| 77 |
+
"mtp_loss": 0.5291013084352016
|
| 78 |
+
},
|
| 79 |
+
{
|
| 80 |
+
"secs": 60,
|
| 81 |
+
"step": 372,
|
| 82 |
+
"train_loss": 0.11918953880667686,
|
| 83 |
+
"val_loss": 0.10963512398302555,
|
| 84 |
+
"mtp_loss": 0.11610854640603066
|
| 85 |
+
},
|
| 86 |
+
{
|
| 87 |
+
"secs": 75,
|
| 88 |
+
"step": 563,
|
| 89 |
+
"train_loss": 0.09439869225025177,
|
| 90 |
+
"val_loss": 0.08949651103466749,
|
| 91 |
+
"mtp_loss": 0.09490611925721168
|
| 92 |
+
},
|
| 93 |
+
{
|
| 94 |
+
"secs": 90,
|
| 95 |
+
"step": 762,
|
| 96 |
+
"train_loss": 0.08308708146214486,
|
| 97 |
+
"val_loss": 0.09794155973941088,
|
| 98 |
+
"mtp_loss": 0.08327573388814927
|
| 99 |
+
},
|
| 100 |
+
{
|
| 101 |
+
"secs": 105,
|
| 102 |
+
"step": 952,
|
| 103 |
+
"train_loss": 0.0701075332860152,
|
| 104 |
+
"val_loss": 0.09880032576620579,
|
| 105 |
+
"mtp_loss": 0.07315145639909638
|
| 106 |
+
},
|
| 107 |
+
{
|
| 108 |
+
"secs": 120,
|
| 109 |
+
"step": 1148,
|
| 110 |
+
"train_loss": 0.07572940364480019,
|
| 111 |
+
"val_loss": 0.08534885942935944,
|
| 112 |
+
"mtp_loss": 0.07808437496423722
|
| 113 |
+
},
|
| 114 |
+
{
|
| 115 |
+
"secs": 135,
|
| 116 |
+
"step": 1342,
|
| 117 |
+
"train_loss": 0.0756804671138525,
|
| 118 |
+
"val_loss": 0.08864960819482803,
|
| 119 |
+
"mtp_loss": 0.07646373771131039
|
| 120 |
+
},
|
| 121 |
+
{
|
| 122 |
+
"secs": 150,
|
| 123 |
+
"step": 1538,
|
| 124 |
+
"train_loss": 0.0630389153957367,
|
| 125 |
+
"val_loss": 0.09432582464069128,
|
| 126 |
+
"mtp_loss": 0.061678790383868747
|
| 127 |
+
},
|
| 128 |
+
{
|
| 129 |
+
"secs": 165,
|
| 130 |
+
"step": 1739,
|
| 131 |
+
"train_loss": 0.05705111175775528,
|
| 132 |
+
"val_loss": 0.08685819059610367,
|
| 133 |
+
"mtp_loss": 0.05723906978964806
|
| 134 |
+
},
|
| 135 |
+
{
|
| 136 |
+
"secs": 180,
|
| 137 |
+
"step": 1933,
|
| 138 |
+
"train_loss": 0.0693660531193018,
|
| 139 |
+
"val_loss": 0.07990295719355345,
|
| 140 |
+
"mtp_loss": 0.0706000704318285
|
| 141 |
+
},
|
| 142 |
+
{
|
| 143 |
+
"secs": 195,
|
| 144 |
+
"step": 2118,
|
| 145 |
+
"train_loss": 0.05955390425191985,
|
| 146 |
+
"val_loss": 0.08126363344490528,
|
| 147 |
+
"mtp_loss": 0.06039786752727297
|
| 148 |
+
},
|
| 149 |
+
{
|
| 150 |
+
"secs": 210,
|
| 151 |
+
"step": 2313,
|
| 152 |
+
"train_loss": 0.05665931235998869,
|
| 153 |
+
"val_loss": 0.08018658589571714,
|
| 154 |
+
"mtp_loss": 0.05772969778627157
|
| 155 |
+
},
|
| 156 |
+
{
|
| 157 |
+
"secs": 225,
|
| 158 |
+
"step": 2510,
|
| 159 |
+
"train_loss": 0.06889263167977333,
|
| 160 |
+
"val_loss": 0.08516997005790472,
|
| 161 |
+
"mtp_loss": 0.06859188210219144
|
| 162 |
+
},
|
| 163 |
+
{
|
| 164 |
+
"secs": 240,
|
| 165 |
+
"step": 2703,
|
| 166 |
+
"train_loss": 0.05151104256510734,
|
| 167 |
+
"val_loss": 0.08512604422867298,
|
| 168 |
+
"mtp_loss": 0.05167849510908127
|
| 169 |
+
},
|
| 170 |
+
{
|
| 171 |
+
"secs": 255,
|
| 172 |
+
"step": 2896,
|
| 173 |
+
"train_loss": 0.05312770770655738,
|
| 174 |
+
"val_loss": 0.08225905895233154,
|
| 175 |
+
"mtp_loss": 0.05338138383295801
|
| 176 |
+
},
|
| 177 |
+
{
|
| 178 |
+
"secs": 270,
|
| 179 |
+
"step": 3081,
|
| 180 |
+
"train_loss": 0.0616743054240942,
|
| 181 |
+
"val_loss": 0.0793133769184351,
|
| 182 |
+
"mtp_loss": 0.06108860932290554
|
| 183 |
+
},
|
| 184 |
+
{
|
| 185 |
+
"secs": 285,
|
| 186 |
+
"step": 3276,
|
| 187 |
+
"train_loss": 0.05170489785571893,
|
| 188 |
+
"val_loss": 0.08151316083967686,
|
| 189 |
+
"mtp_loss": 0.051792557454771466
|
| 190 |
+
},
|
| 191 |
+
{
|
| 192 |
+
"secs": 300,
|
| 193 |
+
"step": 3470,
|
| 194 |
+
"train_loss": 0.05720296762883663,
|
| 195 |
+
"val_loss": 0.08942204620689154,
|
| 196 |
+
"mtp_loss": 0.05792026892304421
|
| 197 |
+
}
|
| 198 |
+
],
|
| 199 |
+
"secs": 301.0619282722473,
|
| 200 |
+
"n_params": 6400256,
|
| 201 |
+
"heldout_greedy": 0.96
|
| 202 |
+
},
|
| 203 |
+
"rl": {
|
| 204 |
+
"rounds": [
|
| 205 |
+
{
|
| 206 |
+
"round": 1,
|
| 207 |
+
"level": 1,
|
| 208 |
+
"k": 8,
|
| 209 |
+
"n_targets": 1500,
|
| 210 |
+
"sample_acc": 0.11141666666666666,
|
| 211 |
+
"new_proofs": 461,
|
| 212 |
+
"targets_solved": 406,
|
| 213 |
+
"lstar": 9,
|
| 214 |
+
"ge": {
|
| 215 |
+
"7": 406,
|
| 216 |
+
"8": 127,
|
| 217 |
+
"9": 17,
|
| 218 |
+
"10": 0,
|
| 219 |
+
"11": 0,
|
| 220 |
+
"12": 0,
|
| 221 |
+
"13": 0,
|
| 222 |
+
"14": 0
|
| 223 |
+
},
|
| 224 |
+
"mix_rl_records": 461,
|
| 225 |
+
"sampling": {
|
| 226 |
+
"samples": 12000,
|
| 227 |
+
"secs": 51.690961837768555,
|
| 228 |
+
"gate_secs": 23.85087752342224,
|
| 229 |
+
"samples_per_sec": 232.14890134298238
|
| 230 |
+
},
|
| 231 |
+
"secs": 52.07204484939575,
|
| 232 |
+
"rl_secs_used": 28.221420764923096
|
| 233 |
+
},
|
| 234 |
+
{
|
| 235 |
+
"round": 2,
|
| 236 |
+
"level": 2,
|
| 237 |
+
"k": 24,
|
| 238 |
+
"n_targets": 365,
|
| 239 |
+
"sample_acc": 0.06472602739726027,
|
| 240 |
+
"new_proofs": 115,
|
| 241 |
+
"targets_solved": 498,
|
| 242 |
+
"lstar": 9,
|
| 243 |
+
"ge": {
|
| 244 |
+
"7": 498,
|
| 245 |
+
"8": 176,
|
| 246 |
+
"9": 38,
|
| 247 |
+
"10": 3,
|
| 248 |
+
"11": 0,
|
| 249 |
+
"12": 0,
|
| 250 |
+
"13": 0,
|
| 251 |
+
"14": 0
|
| 252 |
+
},
|
| 253 |
+
"mix_rl_records": 115,
|
| 254 |
+
"sampling": {
|
| 255 |
+
"samples": 8760,
|
| 256 |
+
"secs": 44.677098751068115,
|
| 257 |
+
"gate_secs": 20.54002094268799,
|
| 258 |
+
"samples_per_sec": 196.0736091841812
|
| 259 |
+
},
|
| 260 |
+
"secs": 44.9638090133667,
|
| 261 |
+
"rl_secs_used": 52.64503812789917
|
| 262 |
+
},
|
| 263 |
+
{
|
| 264 |
+
"round": 3,
|
| 265 |
+
"level": 4,
|
| 266 |
+
"k": 24,
|
| 267 |
+
"n_targets": 365,
|
| 268 |
+
"sample_acc": 0.11073059360730593,
|
| 269 |
+
"new_proofs": 139,
|
| 270 |
+
"targets_solved": 607,
|
| 271 |
+
"lstar": 10,
|
| 272 |
+
"ge": {
|
| 273 |
+
"7": 607,
|
| 274 |
+
"8": 231,
|
| 275 |
+
"9": 62,
|
| 276 |
+
"10": 9,
|
| 277 |
+
"11": 0,
|
| 278 |
+
"12": 0,
|
| 279 |
+
"13": 0,
|
| 280 |
+
"14": 0
|
| 281 |
+
},
|
| 282 |
+
"mix_rl_records": 139,
|
| 283 |
+
"sampling": {
|
| 284 |
+
"samples": 8760,
|
| 285 |
+
"secs": 56.54988479614258,
|
| 286 |
+
"gate_secs": 21.94957184791565,
|
| 287 |
+
"samples_per_sec": 154.90747738176725
|
| 288 |
+
},
|
| 289 |
+
"secs": 56.85590958595276,
|
| 290 |
+
"rl_secs_used": 87.55136609077454
|
| 291 |
+
},
|
| 292 |
+
{
|
| 293 |
+
"round": 4,
|
| 294 |
+
"level": 8,
|
| 295 |
+
"k": 24,
|
| 296 |
+
"n_targets": 364,
|
| 297 |
+
"sample_acc": 0.13038003663003664,
|
| 298 |
+
"new_proofs": 157,
|
| 299 |
+
"targets_solved": 723,
|
| 300 |
+
"lstar": 10,
|
| 301 |
+
"ge": {
|
| 302 |
+
"7": 723,
|
| 303 |
+
"8": 293,
|
| 304 |
+
"9": 96,
|
| 305 |
+
"10": 16,
|
| 306 |
+
"11": 0,
|
| 307 |
+
"12": 0,
|
| 308 |
+
"13": 0,
|
| 309 |
+
"14": 0
|
| 310 |
+
},
|
| 311 |
+
"mix_rl_records": 157,
|
| 312 |
+
"sampling": {
|
| 313 |
+
"samples": 8736,
|
| 314 |
+
"secs": 82.23587322235107,
|
| 315 |
+
"gate_secs": 21.489389657974243,
|
| 316 |
+
"samples_per_sec": 106.23101157300805
|
| 317 |
+
},
|
| 318 |
+
"secs": 82.55112838745117,
|
| 319 |
+
"rl_secs_used": 148.6131236553192
|
| 320 |
+
},
|
| 321 |
+
{
|
| 322 |
+
"round": 5,
|
| 323 |
+
"level": 1,
|
| 324 |
+
"k": 8,
|
| 325 |
+
"n_targets": 1500,
|
| 326 |
+
"sample_acc": 0.4096666666666667,
|
| 327 |
+
"new_proofs": 348,
|
| 328 |
+
"targets_solved": 848,
|
| 329 |
+
"lstar": 10,
|
| 330 |
+
"ge": {
|
| 331 |
+
"7": 848,
|
| 332 |
+
"8": 372,
|
| 333 |
+
"9": 151,
|
| 334 |
+
"10": 31,
|
| 335 |
+
"11": 1,
|
| 336 |
+
"12": 0,
|
| 337 |
+
"13": 0,
|
| 338 |
+
"14": 0
|
| 339 |
+
},
|
| 340 |
+
"mix_rl_records": 1103,
|
| 341 |
+
"sampling": {
|
| 342 |
+
"samples": 12000,
|
| 343 |
+
"secs": 52.27893853187561,
|
| 344 |
+
"gate_secs": 30.770501375198364,
|
| 345 |
+
"samples_per_sec": 229.53794275458247
|
| 346 |
+
},
|
| 347 |
+
"secs": 83.04121017456055,
|
| 348 |
+
"rl_secs_used": 200.88395142555237
|
| 349 |
+
},
|
| 350 |
+
{
|
| 351 |
+
"round": 6,
|
| 352 |
+
"level": 2,
|
| 353 |
+
"k": 24,
|
| 354 |
+
"n_targets": 227,
|
| 355 |
+
"sample_acc": 0.045337738619676944,
|
| 356 |
+
"new_proofs": 40,
|
| 357 |
+
"targets_solved": 875,
|
| 358 |
+
"lstar": 10,
|
| 359 |
+
"ge": {
|
| 360 |
+
"7": 875,
|
| 361 |
+
"8": 390,
|
| 362 |
+
"9": 165,
|
| 363 |
+
"10": 34,
|
| 364 |
+
"11": 1,
|
| 365 |
+
"12": 0,
|
| 366 |
+
"13": 0,
|
| 367 |
+
"14": 0
|
| 368 |
+
},
|
| 369 |
+
"mix_rl_records": 49,
|
| 370 |
+
"sampling": {
|
| 371 |
+
"samples": 5448,
|
| 372 |
+
"secs": 30.638378143310547,
|
| 373 |
+
"gate_secs": 16.321690797805786,
|
| 374 |
+
"samples_per_sec": 177.81620079617343
|
| 375 |
+
},
|
| 376 |
+
"secs": 30.850301027297974,
|
| 377 |
+
"rl_secs_used": 215.41240119934082
|
| 378 |
+
},
|
| 379 |
+
{
|
| 380 |
+
"round": 7,
|
| 381 |
+
"level": 4,
|
| 382 |
+
"k": 24,
|
| 383 |
+
"n_targets": 227,
|
| 384 |
+
"sample_acc": 0.09508076358296623,
|
| 385 |
+
"new_proofs": 85,
|
| 386 |
+
"targets_solved": 930,
|
| 387 |
+
"lstar": 10,
|
| 388 |
+
"ge": {
|
| 389 |
+
"7": 930,
|
| 390 |
+
"8": 431,
|
| 391 |
+
"9": 191,
|
| 392 |
+
"10": 41,
|
| 393 |
+
"11": 2,
|
| 394 |
+
"12": 0,
|
| 395 |
+
"13": 0,
|
| 396 |
+
"14": 0
|
| 397 |
+
},
|
| 398 |
+
"mix_rl_records": 93,
|
| 399 |
+
"sampling": {
|
| 400 |
+
"samples": 5448,
|
| 401 |
+
"secs": 38.43360733985901,
|
| 402 |
+
"gate_secs": 16.763285398483276,
|
| 403 |
+
"samples_per_sec": 141.7509408321906
|
| 404 |
+
},
|
| 405 |
+
"secs": 38.7121946811676,
|
| 406 |
+
"rl_secs_used": 237.36132144927979
|
| 407 |
+
},
|
| 408 |
+
{
|
| 409 |
+
"round": 8,
|
| 410 |
+
"level": 8,
|
| 411 |
+
"k": 24,
|
| 412 |
+
"n_targets": 226,
|
| 413 |
+
"sample_acc": 0.12168141592920353,
|
| 414 |
+
"new_proofs": 91,
|
| 415 |
+
"targets_solved": 988,
|
| 416 |
+
"lstar": 10,
|
| 417 |
+
"ge": {
|
| 418 |
+
"7": 988,
|
| 419 |
+
"8": 469,
|
| 420 |
+
"9": 218,
|
| 421 |
+
"10": 45,
|
| 422 |
+
"11": 4,
|
| 423 |
+
"12": 0,
|
| 424 |
+
"13": 0,
|
| 425 |
+
"14": 0
|
| 426 |
+
},
|
| 427 |
+
"mix_rl_records": 102,
|
| 428 |
+
"sampling": {
|
| 429 |
+
"samples": 5424,
|
| 430 |
+
"secs": 48.45755863189697,
|
| 431 |
+
"gate_secs": 16.037729263305664,
|
| 432 |
+
"samples_per_sec": 111.93300185019383
|
| 433 |
+
},
|
| 434 |
+
"secs": 48.69038152694702,
|
| 435 |
+
"rl_secs_used": 270.01398611068726
|
| 436 |
+
},
|
| 437 |
+
{
|
| 438 |
+
"round": 9,
|
| 439 |
+
"level": 1,
|
| 440 |
+
"k": 8,
|
| 441 |
+
"n_targets": 1500,
|
| 442 |
+
"sample_acc": 0.49575,
|
| 443 |
+
"new_proofs": 305,
|
| 444 |
+
"targets_solved": 1026,
|
| 445 |
+
"lstar": 11,
|
| 446 |
+
"ge": {
|
| 447 |
+
"7": 1026,
|
| 448 |
+
"8": 499,
|
| 449 |
+
"9": 241,
|
| 450 |
+
"10": 52,
|
| 451 |
+
"11": 5,
|
| 452 |
+
"12": 0,
|
| 453 |
+
"13": 0,
|
| 454 |
+
"14": 0
|
| 455 |
+
},
|
| 456 |
+
"mix_rl_records": 1417,
|
| 457 |
+
"sampling": {
|
| 458 |
+
"samples": 12000,
|
| 459 |
+
"secs": 51.44225001335144,
|
| 460 |
+
"gate_secs": 27.655930757522583,
|
| 461 |
+
"samples_per_sec": 233.2712895895007
|
| 462 |
+
},
|
| 463 |
+
"secs": 84.39623880386353,
|
| 464 |
+
"rl_secs_used": 326.7545464038849
|
| 465 |
+
},
|
| 466 |
+
{
|
| 467 |
+
"round": 10,
|
| 468 |
+
"level": 2,
|
| 469 |
+
"k": 24,
|
| 470 |
+
"n_targets": 181,
|
| 471 |
+
"sample_acc": 0.072744014732965,
|
| 472 |
+
"new_proofs": 36,
|
| 473 |
+
"targets_solved": 1041,
|
| 474 |
+
"lstar": 11,
|
| 475 |
+
"ge": {
|
| 476 |
+
"7": 1041,
|
| 477 |
+
"8": 512,
|
| 478 |
+
"9": 250,
|
| 479 |
+
"10": 56,
|
| 480 |
+
"11": 5,
|
| 481 |
+
"12": 0,
|
| 482 |
+
"13": 0,
|
| 483 |
+
"14": 0
|
| 484 |
+
},
|
| 485 |
+
"mix_rl_records": 59,
|
| 486 |
+
"sampling": {
|
| 487 |
+
"samples": 4344,
|
| 488 |
+
"secs": 24.432981252670288,
|
| 489 |
+
"gate_secs": 11.687873601913452,
|
| 490 |
+
"samples_per_sec": 177.7924664647808
|
| 491 |
+
},
|
| 492 |
+
"secs": 24.614425659179688,
|
| 493 |
+
"rl_secs_used": 339.6808228492737
|
| 494 |
+
},
|
| 495 |
+
{
|
| 496 |
+
"round": 11,
|
| 497 |
+
"level": 4,
|
| 498 |
+
"k": 24,
|
| 499 |
+
"n_targets": 181,
|
| 500 |
+
"sample_acc": 0.10405156537753223,
|
| 501 |
+
"new_proofs": 63,
|
| 502 |
+
"targets_solved": 1066,
|
| 503 |
+
"lstar": 11,
|
| 504 |
+
"ge": {
|
| 505 |
+
"7": 1066,
|
| 506 |
+
"8": 530,
|
| 507 |
+
"9": 262,
|
| 508 |
+
"10": 59,
|
| 509 |
+
"11": 6,
|
| 510 |
+
"12": 0,
|
| 511 |
+
"13": 0,
|
| 512 |
+
"14": 0
|
| 513 |
+
},
|
| 514 |
+
"mix_rl_records": 89,
|
| 515 |
+
"sampling": {
|
| 516 |
+
"samples": 4344,
|
| 517 |
+
"secs": 28.863463163375854,
|
| 518 |
+
"gate_secs": 8.289448261260986,
|
| 519 |
+
"samples_per_sec": 150.50169050788043
|
| 520 |
+
},
|
| 521 |
+
"secs": 29.040106296539307,
|
| 522 |
+
"rl_secs_used": 360.43148136138916
|
| 523 |
+
},
|
| 524 |
+
{
|
| 525 |
+
"round": 12,
|
| 526 |
+
"level": 8,
|
| 527 |
+
"k": 24,
|
| 528 |
+
"n_targets": 180,
|
| 529 |
+
"sample_acc": 0.12384259259259259,
|
| 530 |
+
"new_proofs": 55,
|
| 531 |
+
"targets_solved": 1089,
|
| 532 |
+
"lstar": 11,
|
| 533 |
+
"ge": {
|
| 534 |
+
"7": 1089,
|
| 535 |
+
"8": 549,
|
| 536 |
+
"9": 276,
|
| 537 |
+
"10": 62,
|
| 538 |
+
"11": 8,
|
| 539 |
+
"12": 0,
|
| 540 |
+
"13": 0,
|
| 541 |
+
"14": 0
|
| 542 |
+
},
|
| 543 |
+
"mix_rl_records": 87,
|
| 544 |
+
"sampling": {
|
| 545 |
+
"samples": 4320,
|
| 546 |
+
"secs": 39.94521427154541,
|
| 547 |
+
"gate_secs": 9.331846714019775,
|
| 548 |
+
"samples_per_sec": 108.14812434433004
|
| 549 |
+
},
|
| 550 |
+
"secs": 40.12682032585144,
|
| 551 |
+
"rl_secs_used": 391.2264475822449
|
| 552 |
+
},
|
| 553 |
+
{
|
| 554 |
+
"round": 13,
|
| 555 |
+
"level": 1,
|
| 556 |
+
"k": 8,
|
| 557 |
+
"n_targets": 1500,
|
| 558 |
+
"sample_acc": 0.5411666666666667,
|
| 559 |
+
"new_proofs": 394,
|
| 560 |
+
"targets_solved": 1119,
|
| 561 |
+
"lstar": 11,
|
| 562 |
+
"ge": {
|
| 563 |
+
"7": 1119,
|
| 564 |
+
"8": 574,
|
| 565 |
+
"9": 295,
|
| 566 |
+
"10": 68,
|
| 567 |
+
"11": 9,
|
| 568 |
+
"12": 0,
|
| 569 |
+
"13": 0,
|
| 570 |
+
"14": 0
|
| 571 |
+
},
|
| 572 |
+
"mix_rl_records": 1803,
|
| 573 |
+
"sampling": {
|
| 574 |
+
"samples": 12000,
|
| 575 |
+
"secs": 52.059730529785156,
|
| 576 |
+
"gate_secs": 28.80413007736206,
|
| 577 |
+
"samples_per_sec": 230.50445858789817
|
| 578 |
+
},
|
| 579 |
+
"secs": 88.00629234313965,
|
| 580 |
+
"rl_secs_used": 450.4288716316223
|
| 581 |
+
},
|
| 582 |
+
{
|
| 583 |
+
"round": 14,
|
| 584 |
+
"level": 2,
|
| 585 |
+
"k": 24,
|
| 586 |
+
"n_targets": 147,
|
| 587 |
+
"sample_acc": 0.06660997732426303,
|
| 588 |
+
"new_proofs": 34,
|
| 589 |
+
"targets_solved": 1133,
|
| 590 |
+
"lstar": 11,
|
| 591 |
+
"ge": {
|
| 592 |
+
"7": 1133,
|
| 593 |
+
"8": 585,
|
| 594 |
+
"9": 304,
|
| 595 |
+
"10": 71,
|
| 596 |
+
"11": 11,
|
| 597 |
+
"12": 1,
|
| 598 |
+
"13": 0,
|
| 599 |
+
"14": 0
|
| 600 |
+
},
|
| 601 |
+
"mix_rl_records": 50,
|
| 602 |
+
"sampling": {
|
| 603 |
+
"samples": 3528,
|
| 604 |
+
"secs": 17.734636783599854,
|
| 605 |
+
"gate_secs": 6.626576662063599,
|
| 606 |
+
"samples_per_sec": 198.93274630030913
|
| 607 |
+
},
|
| 608 |
+
"secs": 17.87173819541931,
|
| 609 |
+
"rl_secs_used": 461.6737620830536
|
| 610 |
+
},
|
| 611 |
+
{
|
| 612 |
+
"round": 15,
|
| 613 |
+
"level": 4,
|
| 614 |
+
"k": 24,
|
| 615 |
+
"n_targets": 147,
|
| 616 |
+
"sample_acc": 0.08134920634920635,
|
| 617 |
+
"new_proofs": 32,
|
| 618 |
+
"targets_solved": 1148,
|
| 619 |
+
"lstar": 11,
|
| 620 |
+
"ge": {
|
| 621 |
+
"7": 1148,
|
| 622 |
+
"8": 595,
|
| 623 |
+
"9": 313,
|
| 624 |
+
"10": 74,
|
| 625 |
+
"11": 12,
|
| 626 |
+
"12": 1,
|
| 627 |
+
"13": 0,
|
| 628 |
+
"14": 0
|
| 629 |
+
},
|
| 630 |
+
"mix_rl_records": 54,
|
| 631 |
+
"sampling": {
|
| 632 |
+
"samples": 3528,
|
| 633 |
+
"secs": 22.63219141960144,
|
| 634 |
+
"gate_secs": 7.714325428009033,
|
| 635 |
+
"samples_per_sec": 155.88415344280122
|
| 636 |
+
},
|
| 637 |
+
"secs": 22.774674892425537,
|
| 638 |
+
"rl_secs_used": 476.73413372039795
|
| 639 |
+
},
|
| 640 |
+
{
|
| 641 |
+
"round": 16,
|
| 642 |
+
"level": 8,
|
| 643 |
+
"k": 24,
|
| 644 |
+
"n_targets": 146,
|
| 645 |
+
"sample_acc": 0.15439497716894976,
|
| 646 |
+
"new_proofs": 40,
|
| 647 |
+
"targets_solved": 1168,
|
| 648 |
+
"lstar": 11,
|
| 649 |
+
"ge": {
|
| 650 |
+
"7": 1168,
|
| 651 |
+
"8": 611,
|
| 652 |
+
"9": 324,
|
| 653 |
+
"10": 77,
|
| 654 |
+
"11": 14,
|
| 655 |
+
"12": 2,
|
| 656 |
+
"13": 0,
|
| 657 |
+
"14": 0
|
| 658 |
+
},
|
| 659 |
+
"mix_rl_records": 79,
|
| 660 |
+
"sampling": {
|
| 661 |
+
"samples": 3504,
|
| 662 |
+
"secs": 44.52782940864563,
|
| 663 |
+
"gate_secs": 8.051237106323242,
|
| 664 |
+
"samples_per_sec": 78.6923604077511
|
| 665 |
+
},
|
| 666 |
+
"secs": 44.672483682632446,
|
| 667 |
+
"rl_secs_used": 513.3553586006165
|
| 668 |
+
},
|
| 669 |
+
{
|
| 670 |
+
"round": 17,
|
| 671 |
+
"level": 1,
|
| 672 |
+
"k": 8,
|
| 673 |
+
"n_targets": 1500,
|
| 674 |
+
"sample_acc": 0.5786666666666667,
|
| 675 |
+
"new_proofs": 278,
|
| 676 |
+
"targets_solved": 1181,
|
| 677 |
+
"lstar": 11,
|
| 678 |
+
"ge": {
|
| 679 |
+
"7": 1181,
|
| 680 |
+
"8": 622,
|
| 681 |
+
"9": 331,
|
| 682 |
+
"10": 82,
|
| 683 |
+
"11": 15,
|
| 684 |
+
"12": 2,
|
| 685 |
+
"13": 0,
|
| 686 |
+
"14": 0
|
| 687 |
+
},
|
| 688 |
+
"mix_rl_records": 1870,
|
| 689 |
+
"sampling": {
|
| 690 |
+
"samples": 12000,
|
| 691 |
+
"secs": 55.340720891952515,
|
| 692 |
+
"gate_secs": 31.606099128723145,
|
| 693 |
+
"samples_per_sec": 216.8385197480325
|
| 694 |
+
},
|
| 695 |
+
"secs": 91.60855627059937,
|
| 696 |
+
"rl_secs_used": 573.3582246303558
|
| 697 |
+
},
|
| 698 |
+
{
|
| 699 |
+
"round": 18,
|
| 700 |
+
"level": 2,
|
| 701 |
+
"k": 24,
|
| 702 |
+
"n_targets": 131,
|
| 703 |
+
"sample_acc": 0.06743002544529263,
|
| 704 |
+
"new_proofs": 17,
|
| 705 |
+
"targets_solved": 1188,
|
| 706 |
+
"lstar": 11,
|
| 707 |
+
"ge": {
|
| 708 |
+
"7": 1188,
|
| 709 |
+
"8": 628,
|
| 710 |
+
"9": 337,
|
| 711 |
+
"10": 85,
|
| 712 |
+
"11": 16,
|
| 713 |
+
"12": 2,
|
| 714 |
+
"13": 0,
|
| 715 |
+
"14": 0
|
| 716 |
+
},
|
| 717 |
+
"mix_rl_records": 47,
|
| 718 |
+
"sampling": {
|
| 719 |
+
"samples": 3144,
|
| 720 |
+
"secs": 19.443903923034668,
|
| 721 |
+
"gate_secs": 5.588464260101318,
|
| 722 |
+
"samples_per_sec": 161.6959234341509
|
| 723 |
+
},
|
| 724 |
+
"secs": 19.5654296875,
|
| 725 |
+
"rl_secs_used": 587.3347780704498
|
| 726 |
+
},
|
| 727 |
+
{
|
| 728 |
+
"round": 19,
|
| 729 |
+
"level": 4,
|
| 730 |
+
"k": 24,
|
| 731 |
+
"n_targets": 130,
|
| 732 |
+
"sample_acc": 0.12532051282051282,
|
| 733 |
+
"new_proofs": 30,
|
| 734 |
+
"targets_solved": 1200,
|
| 735 |
+
"lstar": 11,
|
| 736 |
+
"ge": {
|
| 737 |
+
"7": 1200,
|
| 738 |
+
"8": 639,
|
| 739 |
+
"9": 345,
|
| 740 |
+
"10": 91,
|
| 741 |
+
"11": 17,
|
| 742 |
+
"12": 2,
|
| 743 |
+
"13": 0,
|
| 744 |
+
"14": 0
|
| 745 |
+
},
|
| 746 |
+
"mix_rl_records": 67,
|
| 747 |
+
"sampling": {
|
| 748 |
+
"samples": 3120,
|
| 749 |
+
"secs": 25.78796911239624,
|
| 750 |
+
"gate_secs": 7.732923269271851,
|
| 751 |
+
"samples_per_sec": 120.9866502632121
|
| 752 |
+
},
|
| 753 |
+
"secs": 25.910351753234863,
|
| 754 |
+
"rl_secs_used": 605.5122120380402
|
| 755 |
+
},
|
| 756 |
+
{
|
| 757 |
+
"round": 20,
|
| 758 |
+
"level": 8,
|
| 759 |
+
"k": 24,
|
| 760 |
+
"n_targets": 130,
|
| 761 |
+
"sample_acc": 0.12211538461538461,
|
| 762 |
+
"new_proofs": 33,
|
| 763 |
+
"targets_solved": 1214,
|
| 764 |
+
"lstar": 11,
|
| 765 |
+
"ge": {
|
| 766 |
+
"7": 1214,
|
| 767 |
+
"8": 653,
|
| 768 |
+
"9": 354,
|
| 769 |
+
"10": 95,
|
| 770 |
+
"11": 17,
|
| 771 |
+
"12": 2,
|
| 772 |
+
"13": 0,
|
| 773 |
+
"14": 0
|
| 774 |
+
},
|
| 775 |
+
"mix_rl_records": 57,
|
| 776 |
+
"sampling": {
|
| 777 |
+
"samples": 3120,
|
| 778 |
+
"secs": 38.282958030700684,
|
| 779 |
+
"gate_secs": 7.539533853530884,
|
| 780 |
+
"samples_per_sec": 81.49840452500936
|
| 781 |
+
},
|
| 782 |
+
"secs": 38.40138840675354,
|
| 783 |
+
"rl_secs_used": 636.3740680217743
|
| 784 |
+
},
|
| 785 |
+
{
|
| 786 |
+
"round": 21,
|
| 787 |
+
"level": 1,
|
| 788 |
+
"k": 8,
|
| 789 |
+
"n_targets": 1500,
|
| 790 |
+
"sample_acc": 0.59925,
|
| 791 |
+
"new_proofs": 360,
|
| 792 |
+
"targets_solved": 1226,
|
| 793 |
+
"lstar": 11,
|
| 794 |
+
"ge": {
|
| 795 |
+
"7": 1226,
|
| 796 |
+
"8": 662,
|
| 797 |
+
"9": 361,
|
| 798 |
+
"10": 98,
|
| 799 |
+
"11": 17,
|
| 800 |
+
"12": 2,
|
| 801 |
+
"13": 0,
|
| 802 |
+
"14": 0
|
| 803 |
+
},
|
| 804 |
+
"mix_rl_records": 2149,
|
| 805 |
+
"sampling": {
|
| 806 |
+
"samples": 12000,
|
| 807 |
+
"secs": 45.83187818527222,
|
| 808 |
+
"gate_secs": 24.70815873146057,
|
| 809 |
+
"samples_per_sec": 261.82649446507133
|
| 810 |
+
},
|
| 811 |
+
"secs": 83.43655586242676,
|
| 812 |
+
"rl_secs_used": 695.1026825904846
|
| 813 |
+
},
|
| 814 |
+
{
|
| 815 |
+
"round": 22,
|
| 816 |
+
"level": 2,
|
| 817 |
+
"k": 24,
|
| 818 |
+
"n_targets": 110,
|
| 819 |
+
"sample_acc": 0.07007575757575757,
|
| 820 |
+
"new_proofs": 24,
|
| 821 |
+
"targets_solved": 1236,
|
| 822 |
+
"lstar": 11,
|
| 823 |
+
"ge": {
|
| 824 |
+
"7": 1236,
|
| 825 |
+
"8": 667,
|
| 826 |
+
"9": 364,
|
| 827 |
+
"10": 100,
|
| 828 |
+
"11": 17,
|
| 829 |
+
"12": 2,
|
| 830 |
+
"13": 0,
|
| 831 |
+
"14": 0
|
| 832 |
+
},
|
| 833 |
+
"mix_rl_records": 49,
|
| 834 |
+
"sampling": {
|
| 835 |
+
"samples": 2640,
|
| 836 |
+
"secs": 15.680938482284546,
|
| 837 |
+
"gate_secs": 4.4924187660217285,
|
| 838 |
+
"samples_per_sec": 168.3572703880272
|
| 839 |
+
},
|
| 840 |
+
"secs": 15.782829523086548,
|
| 841 |
+
"rl_secs_used": 706.3928732872009
|
| 842 |
+
},
|
| 843 |
+
{
|
| 844 |
+
"round": 23,
|
| 845 |
+
"level": 4,
|
| 846 |
+
"k": 24,
|
| 847 |
+
"n_targets": 110,
|
| 848 |
+
"sample_acc": 0.11856060606060606,
|
| 849 |
+
"new_proofs": 21,
|
| 850 |
+
"targets_solved": 1246,
|
| 851 |
+
"lstar": 11,
|
| 852 |
+
"ge": {
|
| 853 |
+
"7": 1246,
|
| 854 |
+
"8": 673,
|
| 855 |
+
"9": 369,
|
| 856 |
+
"10": 104,
|
| 857 |
+
"11": 19,
|
| 858 |
+
"12": 2,
|
| 859 |
+
"13": 0,
|
| 860 |
+
"14": 0
|
| 861 |
+
},
|
| 862 |
+
"mix_rl_records": 47,
|
| 863 |
+
"sampling": {
|
| 864 |
+
"samples": 2640,
|
| 865 |
+
"secs": 15.175414562225342,
|
| 866 |
+
"gate_secs": 4.714868783950806,
|
| 867 |
+
"samples_per_sec": 173.96559343897536
|
| 868 |
+
},
|
| 869 |
+
"secs": 15.280608654022217,
|
| 870 |
+
"rl_secs_used": 716.9586043357849
|
| 871 |
+
},
|
| 872 |
+
{
|
| 873 |
+
"round": 24,
|
| 874 |
+
"level": 8,
|
| 875 |
+
"k": 24,
|
| 876 |
+
"n_targets": 109,
|
| 877 |
+
"sample_acc": 0.11850152905198777,
|
| 878 |
+
"new_proofs": 32,
|
| 879 |
+
"targets_solved": 1256,
|
| 880 |
+
"lstar": 11,
|
| 881 |
+
"ge": {
|
| 882 |
+
"7": 1256,
|
| 883 |
+
"8": 678,
|
| 884 |
+
"9": 373,
|
| 885 |
+
"10": 106,
|
| 886 |
+
"11": 20,
|
| 887 |
+
"12": 2,
|
| 888 |
+
"13": 0,
|
| 889 |
+
"14": 0
|
| 890 |
+
},
|
| 891 |
+
"mix_rl_records": 59,
|
| 892 |
+
"sampling": {
|
| 893 |
+
"samples": 2616,
|
| 894 |
+
"secs": 29.040146827697754,
|
| 895 |
+
"gate_secs": 6.244370937347412,
|
| 896 |
+
"samples_per_sec": 90.08218916802878
|
| 897 |
+
},
|
| 898 |
+
"secs": 29.141696214675903,
|
| 899 |
+
"rl_secs_used": 739.8559339046478
|
| 900 |
+
},
|
| 901 |
+
{
|
| 902 |
+
"round": 25,
|
| 903 |
+
"level": 1,
|
| 904 |
+
"k": 8,
|
| 905 |
+
"n_targets": 1500,
|
| 906 |
+
"sample_acc": 0.6280833333333333,
|
| 907 |
+
"new_proofs": 342,
|
| 908 |
+
"targets_solved": 1259,
|
| 909 |
+
"lstar": 11,
|
| 910 |
+
"ge": {
|
| 911 |
+
"7": 1259,
|
| 912 |
+
"8": 681,
|
| 913 |
+
"9": 374,
|
| 914 |
+
"10": 107,
|
| 915 |
+
"11": 20,
|
| 916 |
+
"12": 2,
|
| 917 |
+
"13": 0,
|
| 918 |
+
"14": 0
|
| 919 |
+
},
|
| 920 |
+
"mix_rl_records": 2327,
|
| 921 |
+
"sampling": {
|
| 922 |
+
"samples": 12000,
|
| 923 |
+
"secs": 54.91544461250305,
|
| 924 |
+
"gate_secs": 30.136982917785645,
|
| 925 |
+
"samples_per_sec": 218.51776098099478
|
| 926 |
+
},
|
| 927 |
+
"secs": 92.2239089012146,
|
| 928 |
+
"rl_secs_used": 801.9432067871094
|
| 929 |
+
},
|
| 930 |
+
{
|
| 931 |
+
"round": 26,
|
| 932 |
+
"level": 2,
|
| 933 |
+
"k": 24,
|
| 934 |
+
"n_targets": 102,
|
| 935 |
+
"sample_acc": 0.10416666666666667,
|
| 936 |
+
"new_proofs": 17,
|
| 937 |
+
"targets_solved": 1263,
|
| 938 |
+
"lstar": 11,
|
| 939 |
+
"ge": {
|
| 940 |
+
"7": 1263,
|
| 941 |
+
"8": 684,
|
| 942 |
+
"9": 377,
|
| 943 |
+
"10": 107,
|
| 944 |
+
"11": 20,
|
| 945 |
+
"12": 2,
|
| 946 |
+
"13": 0,
|
| 947 |
+
"14": 0
|
| 948 |
+
},
|
| 949 |
+
"mix_rl_records": 53,
|
| 950 |
+
"sampling": {
|
| 951 |
+
"samples": 2448,
|
| 952 |
+
"secs": 16.353596687316895,
|
| 953 |
+
"gate_secs": 5.976778984069824,
|
| 954 |
+
"samples_per_sec": 149.69184129987488
|
| 955 |
+
},
|
| 956 |
+
"secs": 16.447291612625122,
|
| 957 |
+
"rl_secs_used": 812.4133667945862
|
| 958 |
+
},
|
| 959 |
+
{
|
| 960 |
+
"round": 27,
|
| 961 |
+
"level": 4,
|
| 962 |
+
"k": 24,
|
| 963 |
+
"n_targets": 102,
|
| 964 |
+
"sample_acc": 0.12091503267973856,
|
| 965 |
+
"new_proofs": 18,
|
| 966 |
+
"targets_solved": 1268,
|
| 967 |
+
"lstar": 11,
|
| 968 |
+
"ge": {
|
| 969 |
+
"7": 1268,
|
| 970 |
+
"8": 689,
|
| 971 |
+
"9": 382,
|
| 972 |
+
"10": 108,
|
| 973 |
+
"11": 20,
|
| 974 |
+
"12": 2,
|
| 975 |
+
"13": 0,
|
| 976 |
+
"14": 0
|
| 977 |
+
},
|
| 978 |
+
"mix_rl_records": 38,
|
| 979 |
+
"sampling": {
|
| 980 |
+
"samples": 2448,
|
| 981 |
+
"secs": 22.125916719436646,
|
| 982 |
+
"gate_secs": 5.621021747589111,
|
| 983 |
+
"samples_per_sec": 110.63948359931861
|
| 984 |
+
},
|
| 985 |
+
"secs": 22.215311765670776,
|
| 986 |
+
"rl_secs_used": 829.0076656341553
|
| 987 |
+
},
|
| 988 |
+
{
|
| 989 |
+
"round": 28,
|
| 990 |
+
"level": 8,
|
| 991 |
+
"k": 24,
|
| 992 |
+
"n_targets": 101,
|
| 993 |
+
"sample_acc": 0.12623762376237624,
|
| 994 |
+
"new_proofs": 42,
|
| 995 |
+
"targets_solved": 1275,
|
| 996 |
+
"lstar": 11,
|
| 997 |
+
"ge": {
|
| 998 |
+
"7": 1275,
|
| 999 |
+
"8": 693,
|
| 1000 |
+
"9": 386,
|
| 1001 |
+
"10": 109,
|
| 1002 |
+
"11": 21,
|
| 1003 |
+
"12": 3,
|
| 1004 |
+
"13": 0,
|
| 1005 |
+
"14": 0
|
| 1006 |
+
},
|
| 1007 |
+
"mix_rl_records": 78,
|
| 1008 |
+
"sampling": {
|
| 1009 |
+
"samples": 2424,
|
| 1010 |
+
"secs": 33.48195552825928,
|
| 1011 |
+
"gate_secs": 5.30162239074707,
|
| 1012 |
+
"samples_per_sec": 72.39720505434957
|
| 1013 |
+
},
|
| 1014 |
+
"secs": 33.57149362564087,
|
| 1015 |
+
"rl_secs_used": 857.2775297164917
|
| 1016 |
+
},
|
| 1017 |
+
{
|
| 1018 |
+
"round": 29,
|
| 1019 |
+
"level": 1,
|
| 1020 |
+
"k": 8,
|
| 1021 |
+
"n_targets": 1500,
|
| 1022 |
+
"sample_acc": 0.6104166666666667,
|
| 1023 |
+
"new_proofs": 611,
|
| 1024 |
+
"targets_solved": 1277,
|
| 1025 |
+
"lstar": 11,
|
| 1026 |
+
"ge": {
|
| 1027 |
+
"7": 1277,
|
| 1028 |
+
"8": 693,
|
| 1029 |
+
"9": 386,
|
| 1030 |
+
"10": 109,
|
| 1031 |
+
"11": 21,
|
| 1032 |
+
"12": 3,
|
| 1033 |
+
"13": 0,
|
| 1034 |
+
"14": 0
|
| 1035 |
+
},
|
| 1036 |
+
"mix_rl_records": 2694,
|
| 1037 |
+
"sampling": {
|
| 1038 |
+
"samples": 12000,
|
| 1039 |
+
"secs": 58.913813829422,
|
| 1040 |
+
"gate_secs": 29.1210675239563,
|
| 1041 |
+
"samples_per_sec": 203.6873734697364
|
| 1042 |
+
},
|
| 1043 |
+
"secs": 97.32018065452576,
|
| 1044 |
+
"rl_secs_used": 925.4768824577332
|
| 1045 |
+
},
|
| 1046 |
+
{
|
| 1047 |
+
"round": 30,
|
| 1048 |
+
"level": 2,
|
| 1049 |
+
"k": 24,
|
| 1050 |
+
"n_targets": 95,
|
| 1051 |
+
"sample_acc": 0.12850877192982457,
|
| 1052 |
+
"new_proofs": 24,
|
| 1053 |
+
"targets_solved": 1279,
|
| 1054 |
+
"lstar": 11,
|
| 1055 |
+
"ge": {
|
| 1056 |
+
"7": 1279,
|
| 1057 |
+
"8": 695,
|
| 1058 |
+
"9": 388,
|
| 1059 |
+
"10": 111,
|
| 1060 |
+
"11": 22,
|
| 1061 |
+
"12": 4,
|
| 1062 |
+
"13": 0,
|
| 1063 |
+
"14": 0
|
| 1064 |
+
},
|
| 1065 |
+
"mix_rl_records": 58,
|
| 1066 |
+
"sampling": {
|
| 1067 |
+
"samples": 2280,
|
| 1068 |
+
"secs": 14.684942483901978,
|
| 1069 |
+
"gate_secs": 4.952110052108765,
|
| 1070 |
+
"samples_per_sec": 155.26107797149334
|
| 1071 |
+
},
|
| 1072 |
+
"secs": 14.76914930343628,
|
| 1073 |
+
"rl_secs_used": 935.2936854362488
|
| 1074 |
+
},
|
| 1075 |
+
{
|
| 1076 |
+
"round": 31,
|
| 1077 |
+
"level": 4,
|
| 1078 |
+
"k": 24,
|
| 1079 |
+
"n_targets": 95,
|
| 1080 |
+
"sample_acc": 0.15526315789473685,
|
| 1081 |
+
"new_proofs": 21,
|
| 1082 |
+
"targets_solved": 1283,
|
| 1083 |
+
"lstar": 11,
|
| 1084 |
+
"ge": {
|
| 1085 |
+
"7": 1283,
|
| 1086 |
+
"8": 698,
|
| 1087 |
+
"9": 391,
|
| 1088 |
+
"10": 114,
|
| 1089 |
+
"11": 23,
|
| 1090 |
+
"12": 4,
|
| 1091 |
+
"13": 0,
|
| 1092 |
+
"14": 0
|
| 1093 |
+
},
|
| 1094 |
+
"mix_rl_records": 59,
|
| 1095 |
+
"sampling": {
|
| 1096 |
+
"samples": 2280,
|
| 1097 |
+
"secs": 20.38936448097229,
|
| 1098 |
+
"gate_secs": 5.675371885299683,
|
| 1099 |
+
"samples_per_sec": 111.82300469088851
|
| 1100 |
+
},
|
| 1101 |
+
"secs": 20.48034930229187,
|
| 1102 |
+
"rl_secs_used": 950.0986659526825
|
| 1103 |
+
},
|
| 1104 |
+
{
|
| 1105 |
+
"round": 32,
|
| 1106 |
+
"level": 8,
|
| 1107 |
+
"k": 24,
|
| 1108 |
+
"n_targets": 95,
|
| 1109 |
+
"sample_acc": 0.125,
|
| 1110 |
+
"new_proofs": 15,
|
| 1111 |
+
"targets_solved": 1284,
|
| 1112 |
+
"lstar": 11,
|
| 1113 |
+
"ge": {
|
| 1114 |
+
"7": 1284,
|
| 1115 |
+
"8": 698,
|
| 1116 |
+
"9": 391,
|
| 1117 |
+
"10": 114,
|
| 1118 |
+
"11": 23,
|
| 1119 |
+
"12": 4,
|
| 1120 |
+
"13": 0,
|
| 1121 |
+
"14": 0
|
| 1122 |
+
},
|
| 1123 |
+
"mix_rl_records": 40,
|
| 1124 |
+
"sampling": {
|
| 1125 |
+
"samples": 2280,
|
| 1126 |
+
"secs": 26.46371340751648,
|
| 1127 |
+
"gate_secs": 5.069012641906738,
|
| 1128 |
+
"samples_per_sec": 86.15570932507198
|
| 1129 |
+
},
|
| 1130 |
+
"secs": 26.54768705368042,
|
| 1131 |
+
"rl_secs_used": 971.5773332118988
|
| 1132 |
+
},
|
| 1133 |
+
{
|
| 1134 |
+
"round": 33,
|
| 1135 |
+
"level": 1,
|
| 1136 |
+
"k": 8,
|
| 1137 |
+
"n_targets": 1500,
|
| 1138 |
+
"sample_acc": 0.64925,
|
| 1139 |
+
"new_proofs": 660,
|
| 1140 |
+
"targets_solved": 1285,
|
| 1141 |
+
"lstar": 11,
|
| 1142 |
+
"ge": {
|
| 1143 |
+
"7": 1285,
|
| 1144 |
+
"8": 699,
|
| 1145 |
+
"9": 392,
|
| 1146 |
+
"10": 114,
|
| 1147 |
+
"11": 23,
|
| 1148 |
+
"12": 4,
|
| 1149 |
+
"13": 0,
|
| 1150 |
+
"14": 0
|
| 1151 |
+
},
|
| 1152 |
+
"mix_rl_records": 3037,
|
| 1153 |
+
"sampling": {
|
| 1154 |
+
"samples": 12000,
|
| 1155 |
+
"secs": 56.77757453918457,
|
| 1156 |
+
"gate_secs": 28.348571062088013,
|
| 1157 |
+
"samples_per_sec": 211.35105008260436
|
| 1158 |
+
},
|
| 1159 |
+
"secs": 95.34531354904175,
|
| 1160 |
+
"rl_secs_used": 1038.5744199752808
|
| 1161 |
+
},
|
| 1162 |
+
{
|
| 1163 |
+
"round": 34,
|
| 1164 |
+
"level": 2,
|
| 1165 |
+
"k": 24,
|
| 1166 |
+
"n_targets": 86,
|
| 1167 |
+
"sample_acc": 0.08236434108527131,
|
| 1168 |
+
"new_proofs": 8,
|
| 1169 |
+
"targets_solved": 1286,
|
| 1170 |
+
"lstar": 11,
|
| 1171 |
+
"ge": {
|
| 1172 |
+
"7": 1286,
|
| 1173 |
+
"8": 700,
|
| 1174 |
+
"9": 393,
|
| 1175 |
+
"10": 115,
|
| 1176 |
+
"11": 24,
|
| 1177 |
+
"12": 4,
|
| 1178 |
+
"13": 0,
|
| 1179 |
+
"14": 0
|
| 1180 |
+
},
|
| 1181 |
+
"mix_rl_records": 27,
|
| 1182 |
+
"sampling": {
|
| 1183 |
+
"samples": 2064,
|
| 1184 |
+
"secs": 9.983755588531494,
|
| 1185 |
+
"gate_secs": 2.0450656414031982,
|
| 1186 |
+
"samples_per_sec": 206.73583018908747
|
| 1187 |
+
},
|
| 1188 |
+
"secs": 10.045098304748535,
|
| 1189 |
+
"rl_secs_used": 1046.5741090774536
|
| 1190 |
+
},
|
| 1191 |
+
{
|
| 1192 |
+
"round": 35,
|
| 1193 |
+
"level": 4,
|
| 1194 |
+
"k": 24,
|
| 1195 |
+
"n_targets": 86,
|
| 1196 |
+
"sample_acc": 0.08042635658914729,
|
| 1197 |
+
"new_proofs": 10,
|
| 1198 |
+
"targets_solved": 1289,
|
| 1199 |
+
"lstar": 11,
|
| 1200 |
+
"ge": {
|
| 1201 |
+
"7": 1289,
|
| 1202 |
+
"8": 702,
|
| 1203 |
+
"9": 394,
|
| 1204 |
+
"10": 116,
|
| 1205 |
+
"11": 24,
|
| 1206 |
+
"12": 4,
|
| 1207 |
+
"13": 0,
|
| 1208 |
+
"14": 0
|
| 1209 |
+
},
|
| 1210 |
+
"mix_rl_records": 25,
|
| 1211 |
+
"sampling": {
|
| 1212 |
+
"samples": 2064,
|
| 1213 |
+
"secs": 19.783962965011597,
|
| 1214 |
+
"gate_secs": 4.51764178276062,
|
| 1215 |
+
"samples_per_sec": 104.32692396615543
|
| 1216 |
+
},
|
| 1217 |
+
"secs": 19.88755202293396,
|
| 1218 |
+
"rl_secs_used": 1061.9440324306488
|
| 1219 |
+
},
|
| 1220 |
+
{
|
| 1221 |
+
"round": 36,
|
| 1222 |
+
"level": 8,
|
| 1223 |
+
"k": 24,
|
| 1224 |
+
"n_targets": 85,
|
| 1225 |
+
"sample_acc": 0.10735294117647058,
|
| 1226 |
+
"new_proofs": 25,
|
| 1227 |
+
"targets_solved": 1291,
|
| 1228 |
+
"lstar": 11,
|
| 1229 |
+
"ge": {
|
| 1230 |
+
"7": 1291,
|
| 1231 |
+
"8": 704,
|
| 1232 |
+
"9": 396,
|
| 1233 |
+
"10": 117,
|
| 1234 |
+
"11": 24,
|
| 1235 |
+
"12": 4,
|
| 1236 |
+
"13": 0,
|
| 1237 |
+
"14": 0
|
| 1238 |
+
},
|
| 1239 |
+
"mix_rl_records": 53,
|
| 1240 |
+
"sampling": {
|
| 1241 |
+
"samples": 2040,
|
| 1242 |
+
"secs": 31.45756196975708,
|
| 1243 |
+
"gate_secs": 5.4151482582092285,
|
| 1244 |
+
"samples_per_sec": 64.84927223416841
|
| 1245 |
+
},
|
| 1246 |
+
"secs": 31.534928798675537,
|
| 1247 |
+
"rl_secs_used": 1088.0638010501862
|
| 1248 |
+
},
|
| 1249 |
+
{
|
| 1250 |
+
"round": 37,
|
| 1251 |
+
"level": 1,
|
| 1252 |
+
"k": 8,
|
| 1253 |
+
"n_targets": 1500,
|
| 1254 |
+
"sample_acc": 0.6645,
|
| 1255 |
+
"new_proofs": 584,
|
| 1256 |
+
"targets_solved": 1291,
|
| 1257 |
+
"lstar": 11,
|
| 1258 |
+
"ge": {
|
| 1259 |
+
"7": 1291,
|
| 1260 |
+
"8": 704,
|
| 1261 |
+
"9": 396,
|
| 1262 |
+
"10": 117,
|
| 1263 |
+
"11": 24,
|
| 1264 |
+
"12": 4,
|
| 1265 |
+
"13": 0,
|
| 1266 |
+
"14": 0
|
| 1267 |
+
},
|
| 1268 |
+
"mix_rl_records": 3080,
|
| 1269 |
+
"sampling": {
|
| 1270 |
+
"samples": 12000,
|
| 1271 |
+
"secs": 65.147390127182,
|
| 1272 |
+
"gate_secs": 31.265851974487305,
|
| 1273 |
+
"samples_per_sec": 184.19770886559485
|
| 1274 |
+
},
|
| 1275 |
+
"secs": 104.72338199615479,
|
| 1276 |
+
"rl_secs_used": 1161.5215785503387
|
| 1277 |
+
},
|
| 1278 |
+
{
|
| 1279 |
+
"round": 38,
|
| 1280 |
+
"level": 2,
|
| 1281 |
+
"k": 24,
|
| 1282 |
+
"n_targets": 84,
|
| 1283 |
+
"sample_acc": 0.08978174603174603,
|
| 1284 |
+
"new_proofs": 28,
|
| 1285 |
+
"targets_solved": 1294,
|
| 1286 |
+
"lstar": 11,
|
| 1287 |
+
"ge": {
|
| 1288 |
+
"7": 1294,
|
| 1289 |
+
"8": 705,
|
| 1290 |
+
"9": 397,
|
| 1291 |
+
"10": 118,
|
| 1292 |
+
"11": 24,
|
| 1293 |
+
"12": 4,
|
| 1294 |
+
"13": 0,
|
| 1295 |
+
"14": 0
|
| 1296 |
+
},
|
| 1297 |
+
"mix_rl_records": 63,
|
| 1298 |
+
"sampling": {
|
| 1299 |
+
"samples": 2016,
|
| 1300 |
+
"secs": 12.973832130432129,
|
| 1301 |
+
"gate_secs": 4.695043325424194,
|
| 1302 |
+
"samples_per_sec": 155.38970904912208
|
| 1303 |
+
},
|
| 1304 |
+
"secs": 13.050162076950073,
|
| 1305 |
+
"rl_secs_used": 1169.8764414787292
|
| 1306 |
+
},
|
| 1307 |
+
{
|
| 1308 |
+
"round": 39,
|
| 1309 |
+
"level": 4,
|
| 1310 |
+
"k": 24,
|
| 1311 |
+
"n_targets": 84,
|
| 1312 |
+
"sample_acc": 0.08779761904761904,
|
| 1313 |
+
"new_proofs": 13,
|
| 1314 |
+
"targets_solved": 1294,
|
| 1315 |
+
"lstar": 11,
|
| 1316 |
+
"ge": {
|
| 1317 |
+
"7": 1294,
|
| 1318 |
+
"8": 705,
|
| 1319 |
+
"9": 397,
|
| 1320 |
+
"10": 118,
|
| 1321 |
+
"11": 24,
|
| 1322 |
+
"12": 4,
|
| 1323 |
+
"13": 0,
|
| 1324 |
+
"14": 0
|
| 1325 |
+
},
|
| 1326 |
+
"mix_rl_records": 33,
|
| 1327 |
+
"sampling": {
|
| 1328 |
+
"samples": 2016,
|
| 1329 |
+
"secs": 18.29031491279602,
|
| 1330 |
+
"gate_secs": 4.078568935394287,
|
| 1331 |
+
"samples_per_sec": 110.22226843068698
|
| 1332 |
+
},
|
| 1333 |
+
"secs": 18.35506272315979,
|
| 1334 |
+
"rl_secs_used": 1184.1529369354248
|
| 1335 |
+
},
|
| 1336 |
+
{
|
| 1337 |
+
"round": 40,
|
| 1338 |
+
"level": 8,
|
| 1339 |
+
"k": 24,
|
| 1340 |
+
"n_targets": 83,
|
| 1341 |
+
"sample_acc": 0.0963855421686747,
|
| 1342 |
+
"new_proofs": 18,
|
| 1343 |
+
"targets_solved": 1294,
|
| 1344 |
+
"lstar": 11,
|
| 1345 |
+
"ge": {
|
| 1346 |
+
"7": 1294,
|
| 1347 |
+
"8": 705,
|
| 1348 |
+
"9": 397,
|
| 1349 |
+
"10": 118,
|
| 1350 |
+
"11": 24,
|
| 1351 |
+
"12": 4,
|
| 1352 |
+
"13": 0,
|
| 1353 |
+
"14": 0
|
| 1354 |
+
},
|
| 1355 |
+
"mix_rl_records": 49,
|
| 1356 |
+
"sampling": {
|
| 1357 |
+
"samples": 1992,
|
| 1358 |
+
"secs": 27.630221128463745,
|
| 1359 |
+
"gate_secs": 4.135980129241943,
|
| 1360 |
+
"samples_per_sec": 72.09497132644759
|
| 1361 |
+
},
|
| 1362 |
+
"secs": 27.703002452850342,
|
| 1363 |
+
"rl_secs_used": 1207.7199652194977
|
| 1364 |
+
},
|
| 1365 |
+
{
|
| 1366 |
+
"round": 41,
|
| 1367 |
+
"level": 1,
|
| 1368 |
+
"k": 8,
|
| 1369 |
+
"n_targets": 1500,
|
| 1370 |
+
"sample_acc": 0.6676666666666666,
|
| 1371 |
+
"new_proofs": 930,
|
| 1372 |
+
"targets_solved": 1296,
|
| 1373 |
+
"lstar": 11,
|
| 1374 |
+
"ge": {
|
| 1375 |
+
"7": 1296,
|
| 1376 |
+
"8": 706,
|
| 1377 |
+
"9": 398,
|
| 1378 |
+
"10": 118,
|
| 1379 |
+
"11": 24,
|
| 1380 |
+
"12": 4,
|
| 1381 |
+
"13": 0,
|
| 1382 |
+
"14": 0
|
| 1383 |
+
},
|
| 1384 |
+
"mix_rl_records": 3555,
|
| 1385 |
+
"sampling": {
|
| 1386 |
+
"samples": 12000,
|
| 1387 |
+
"secs": 57.68671274185181,
|
| 1388 |
+
"gate_secs": 26.203401803970337,
|
| 1389 |
+
"samples_per_sec": 208.02017361779707
|
| 1390 |
+
},
|
| 1391 |
+
"secs": 97.84754514694214,
|
| 1392 |
+
"rl_secs_used": 1279.364325761795
|
| 1393 |
+
},
|
| 1394 |
+
{
|
| 1395 |
+
"round": 42,
|
| 1396 |
+
"level": 2,
|
| 1397 |
+
"k": 24,
|
| 1398 |
+
"n_targets": 78,
|
| 1399 |
+
"sample_acc": 0.04487179487179487,
|
| 1400 |
+
"new_proofs": 10,
|
| 1401 |
+
"targets_solved": 1299,
|
| 1402 |
+
"lstar": 11,
|
| 1403 |
+
"ge": {
|
| 1404 |
+
"7": 1299,
|
| 1405 |
+
"8": 707,
|
| 1406 |
+
"9": 398,
|
| 1407 |
+
"10": 118,
|
| 1408 |
+
"11": 24,
|
| 1409 |
+
"12": 4,
|
| 1410 |
+
"13": 0,
|
| 1411 |
+
"14": 0
|
| 1412 |
+
},
|
| 1413 |
+
"mix_rl_records": 26,
|
| 1414 |
+
"sampling": {
|
| 1415 |
+
"samples": 1872,
|
| 1416 |
+
"secs": 11.336847066879272,
|
| 1417 |
+
"gate_secs": 3.547269105911255,
|
| 1418 |
+
"samples_per_sec": 165.12527592165102
|
| 1419 |
+
},
|
| 1420 |
+
"secs": 11.406150579452515,
|
| 1421 |
+
"rl_secs_used": 1287.222981929779
|
| 1422 |
+
},
|
| 1423 |
+
{
|
| 1424 |
+
"round": 43,
|
| 1425 |
+
"level": 4,
|
| 1426 |
+
"k": 24,
|
| 1427 |
+
"n_targets": 78,
|
| 1428 |
+
"sample_acc": 0.05822649572649573,
|
| 1429 |
+
"new_proofs": 18,
|
| 1430 |
+
"targets_solved": 1301,
|
| 1431 |
+
"lstar": 11,
|
| 1432 |
+
"ge": {
|
| 1433 |
+
"7": 1301,
|
| 1434 |
+
"8": 709,
|
| 1435 |
+
"9": 400,
|
| 1436 |
+
"10": 119,
|
| 1437 |
+
"11": 24,
|
| 1438 |
+
"12": 4,
|
| 1439 |
+
"13": 0,
|
| 1440 |
+
"14": 0
|
| 1441 |
+
},
|
| 1442 |
+
"mix_rl_records": 35,
|
| 1443 |
+
"sampling": {
|
| 1444 |
+
"samples": 1872,
|
| 1445 |
+
"secs": 15.453723192214966,
|
| 1446 |
+
"gate_secs": 3.190979242324829,
|
| 1447 |
+
"samples_per_sec": 121.13585682336065
|
| 1448 |
+
},
|
| 1449 |
+
"secs": 15.511287927627563,
|
| 1450 |
+
"rl_secs_used": 1299.5432906150818
|
| 1451 |
+
},
|
| 1452 |
+
{
|
| 1453 |
+
"round": 44,
|
| 1454 |
+
"level": 8,
|
| 1455 |
+
"k": 24,
|
| 1456 |
+
"n_targets": 78,
|
| 1457 |
+
"sample_acc": 0.08707264957264957,
|
| 1458 |
+
"new_proofs": 12,
|
| 1459 |
+
"targets_solved": 1304,
|
| 1460 |
+
"lstar": 11,
|
| 1461 |
+
"ge": {
|
| 1462 |
+
"7": 1304,
|
| 1463 |
+
"8": 711,
|
| 1464 |
+
"9": 402,
|
| 1465 |
+
"10": 120,
|
| 1466 |
+
"11": 25,
|
| 1467 |
+
"12": 4,
|
| 1468 |
+
"13": 0,
|
| 1469 |
+
"14": 0
|
| 1470 |
+
},
|
| 1471 |
+
"mix_rl_records": 34,
|
| 1472 |
+
"sampling": {
|
| 1473 |
+
"samples": 1872,
|
| 1474 |
+
"secs": 26.81679630279541,
|
| 1475 |
+
"gate_secs": 3.738881826400757,
|
| 1476 |
+
"samples_per_sec": 69.80699628929429
|
| 1477 |
+
},
|
| 1478 |
+
"secs": 26.875347137451172,
|
| 1479 |
+
"rl_secs_used": 1322.6797549724579
|
| 1480 |
+
},
|
| 1481 |
+
{
|
| 1482 |
+
"round": 45,
|
| 1483 |
+
"level": 1,
|
| 1484 |
+
"k": 8,
|
| 1485 |
+
"n_targets": 1500,
|
| 1486 |
+
"sample_acc": 0.67075,
|
| 1487 |
+
"new_proofs": 809,
|
| 1488 |
+
"targets_solved": 1304,
|
| 1489 |
+
"lstar": 11,
|
| 1490 |
+
"ge": {
|
| 1491 |
+
"7": 1304,
|
| 1492 |
+
"8": 711,
|
| 1493 |
+
"9": 402,
|
| 1494 |
+
"10": 120,
|
| 1495 |
+
"11": 25,
|
| 1496 |
+
"12": 4,
|
| 1497 |
+
"13": 0,
|
| 1498 |
+
"14": 0
|
| 1499 |
+
},
|
| 1500 |
+
"mix_rl_records": 3654,
|
| 1501 |
+
"sampling": {
|
| 1502 |
+
"samples": 12000,
|
| 1503 |
+
"secs": 59.32733631134033,
|
| 1504 |
+
"gate_secs": 29.665650606155396,
|
| 1505 |
+
"samples_per_sec": 202.26763488968942
|
| 1506 |
+
},
|
| 1507 |
+
"secs": 100.17605423927307,
|
| 1508 |
+
"rl_secs_used": 1393.1904182434082
|
| 1509 |
+
},
|
| 1510 |
+
{
|
| 1511 |
+
"round": 46,
|
| 1512 |
+
"level": 2,
|
| 1513 |
+
"k": 24,
|
| 1514 |
+
"n_targets": 79,
|
| 1515 |
+
"sample_acc": 0.04535864978902954,
|
| 1516 |
+
"new_proofs": 13,
|
| 1517 |
+
"targets_solved": 1304,
|
| 1518 |
+
"lstar": 11,
|
| 1519 |
+
"ge": {
|
| 1520 |
+
"7": 1304,
|
| 1521 |
+
"8": 711,
|
| 1522 |
+
"9": 402,
|
| 1523 |
+
"10": 120,
|
| 1524 |
+
"11": 25,
|
| 1525 |
+
"12": 4,
|
| 1526 |
+
"13": 0,
|
| 1527 |
+
"14": 0
|
| 1528 |
+
},
|
| 1529 |
+
"mix_rl_records": 26,
|
| 1530 |
+
"sampling": {
|
| 1531 |
+
"samples": 1896,
|
| 1532 |
+
"secs": 12.433135271072388,
|
| 1533 |
+
"gate_secs": 4.101438760757446,
|
| 1534 |
+
"samples_per_sec": 152.49572683499528
|
| 1535 |
+
},
|
| 1536 |
+
"secs": 12.499722719192505,
|
| 1537 |
+
"rl_secs_used": 1401.5884456634521
|
| 1538 |
+
},
|
| 1539 |
+
{
|
| 1540 |
+
"round": 47,
|
| 1541 |
+
"level": 4,
|
| 1542 |
+
"k": 24,
|
| 1543 |
+
"n_targets": 78,
|
| 1544 |
+
"sample_acc": 0.08386752136752136,
|
| 1545 |
+
"new_proofs": 17,
|
| 1546 |
+
"targets_solved": 1306,
|
| 1547 |
+
"lstar": 11,
|
| 1548 |
+
"ge": {
|
| 1549 |
+
"7": 1306,
|
| 1550 |
+
"8": 713,
|
| 1551 |
+
"9": 402,
|
| 1552 |
+
"10": 120,
|
| 1553 |
+
"11": 25,
|
| 1554 |
+
"12": 4,
|
| 1555 |
+
"13": 0,
|
| 1556 |
+
"14": 0
|
| 1557 |
+
},
|
| 1558 |
+
"mix_rl_records": 36,
|
| 1559 |
+
"sampling": {
|
| 1560 |
+
"samples": 1872,
|
| 1561 |
+
"secs": 17.192299842834473,
|
| 1562 |
+
"gate_secs": 4.083724021911621,
|
| 1563 |
+
"samples_per_sec": 108.88595575421081
|
| 1564 |
+
},
|
| 1565 |
+
"secs": 17.2546865940094,
|
| 1566 |
+
"rl_secs_used": 1414.759408712387
|
| 1567 |
+
},
|
| 1568 |
+
{
|
| 1569 |
+
"round": 48,
|
| 1570 |
+
"level": 8,
|
| 1571 |
+
"k": 24,
|
| 1572 |
+
"n_targets": 78,
|
| 1573 |
+
"sample_acc": 0.140491452991453,
|
| 1574 |
+
"new_proofs": 32,
|
| 1575 |
+
"targets_solved": 1306,
|
| 1576 |
+
"lstar": 11,
|
| 1577 |
+
"ge": {
|
| 1578 |
+
"7": 1306,
|
| 1579 |
+
"8": 713,
|
| 1580 |
+
"9": 402,
|
| 1581 |
+
"10": 120,
|
| 1582 |
+
"11": 25,
|
| 1583 |
+
"12": 4,
|
| 1584 |
+
"13": 0,
|
| 1585 |
+
"14": 0
|
| 1586 |
+
},
|
| 1587 |
+
"mix_rl_records": 84,
|
| 1588 |
+
"sampling": {
|
| 1589 |
+
"samples": 1872,
|
| 1590 |
+
"secs": 24.505131006240845,
|
| 1591 |
+
"gate_secs": 3.584219455718994,
|
| 1592 |
+
"samples_per_sec": 76.39216454395809
|
| 1593 |
+
},
|
| 1594 |
+
"secs": 24.57925319671631,
|
| 1595 |
+
"rl_secs_used": 1435.7544486522675
|
| 1596 |
+
},
|
| 1597 |
+
{
|
| 1598 |
+
"round": 49,
|
| 1599 |
+
"level": 1,
|
| 1600 |
+
"k": 8,
|
| 1601 |
+
"n_targets": 1500,
|
| 1602 |
+
"sample_acc": 0.6641666666666667,
|
| 1603 |
+
"new_proofs": 1026,
|
| 1604 |
+
"targets_solved": 1306,
|
| 1605 |
+
"lstar": 11,
|
| 1606 |
+
"ge": {
|
| 1607 |
+
"7": 1306,
|
| 1608 |
+
"8": 713,
|
| 1609 |
+
"9": 402,
|
| 1610 |
+
"10": 120,
|
| 1611 |
+
"11": 25,
|
| 1612 |
+
"12": 4,
|
| 1613 |
+
"13": 0,
|
| 1614 |
+
"14": 0
|
| 1615 |
+
},
|
| 1616 |
+
"mix_rl_records": 3948,
|
| 1617 |
+
"sampling": {
|
| 1618 |
+
"samples": 12000,
|
| 1619 |
+
"secs": 63.31440734863281,
|
| 1620 |
+
"gate_secs": 30.213568210601807,
|
| 1621 |
+
"samples_per_sec": 189.53032181006625
|
| 1622 |
+
},
|
| 1623 |
+
"secs": 105.4401957988739,
|
| 1624 |
+
"rl_secs_used": 1510.9813160896301
|
| 1625 |
+
},
|
| 1626 |
+
{
|
| 1627 |
+
"round": 50,
|
| 1628 |
+
"level": 2,
|
| 1629 |
+
"k": 24,
|
| 1630 |
+
"n_targets": 75,
|
| 1631 |
+
"sample_acc": 0.042222222222222223,
|
| 1632 |
+
"new_proofs": 21,
|
| 1633 |
+
"targets_solved": 1308,
|
| 1634 |
+
"lstar": 11,
|
| 1635 |
+
"ge": {
|
| 1636 |
+
"7": 1308,
|
| 1637 |
+
"8": 714,
|
| 1638 |
+
"9": 403,
|
| 1639 |
+
"10": 121,
|
| 1640 |
+
"11": 25,
|
| 1641 |
+
"12": 4,
|
| 1642 |
+
"13": 0,
|
| 1643 |
+
"14": 0
|
| 1644 |
+
},
|
| 1645 |
+
"mix_rl_records": 42,
|
| 1646 |
+
"sampling": {
|
| 1647 |
+
"samples": 1800,
|
| 1648 |
+
"secs": 11.77936577796936,
|
| 1649 |
+
"gate_secs": 3.9211432933807373,
|
| 1650 |
+
"samples_per_sec": 152.80958533153736
|
| 1651 |
+
},
|
| 1652 |
+
"secs": 11.840721607208252,
|
| 1653 |
+
"rl_secs_used": 1518.9006447792053
|
| 1654 |
+
},
|
| 1655 |
+
{
|
| 1656 |
+
"round": 51,
|
| 1657 |
+
"level": 4,
|
| 1658 |
+
"k": 24,
|
| 1659 |
+
"n_targets": 75,
|
| 1660 |
+
"sample_acc": 0.12611111111111112,
|
| 1661 |
+
"new_proofs": 51,
|
| 1662 |
+
"targets_solved": 1310,
|
| 1663 |
+
"lstar": 11,
|
| 1664 |
+
"ge": {
|
| 1665 |
+
"7": 1310,
|
| 1666 |
+
"8": 716,
|
| 1667 |
+
"9": 405,
|
| 1668 |
+
"10": 121,
|
| 1669 |
+
"11": 25,
|
| 1670 |
+
"12": 4,
|
| 1671 |
+
"13": 0,
|
| 1672 |
+
"14": 0
|
| 1673 |
+
},
|
| 1674 |
+
"mix_rl_records": 95,
|
| 1675 |
+
"sampling": {
|
| 1676 |
+
"samples": 1800,
|
| 1677 |
+
"secs": 15.781806945800781,
|
| 1678 |
+
"gate_secs": 3.8871729373931885,
|
| 1679 |
+
"samples_per_sec": 114.05538074199694
|
| 1680 |
+
},
|
| 1681 |
+
"secs": 15.837306261062622,
|
| 1682 |
+
"rl_secs_used": 1530.8507759571075
|
| 1683 |
+
},
|
| 1684 |
+
{
|
| 1685 |
+
"round": 52,
|
| 1686 |
+
"level": 8,
|
| 1687 |
+
"k": 24,
|
| 1688 |
+
"n_targets": 74,
|
| 1689 |
+
"sample_acc": 0.06587837837837837,
|
| 1690 |
+
"new_proofs": 9,
|
| 1691 |
+
"targets_solved": 1311,
|
| 1692 |
+
"lstar": 11,
|
| 1693 |
+
"ge": {
|
| 1694 |
+
"7": 1311,
|
| 1695 |
+
"8": 717,
|
| 1696 |
+
"9": 406,
|
| 1697 |
+
"10": 121,
|
| 1698 |
+
"11": 25,
|
| 1699 |
+
"12": 4,
|
| 1700 |
+
"13": 0,
|
| 1701 |
+
"14": 0
|
| 1702 |
+
},
|
| 1703 |
+
"mix_rl_records": 27,
|
| 1704 |
+
"sampling": {
|
| 1705 |
+
"samples": 1776,
|
| 1706 |
+
"secs": 22.416046380996704,
|
| 1707 |
+
"gate_secs": 2.8855698108673096,
|
| 1708 |
+
"samples_per_sec": 79.22895812285664
|
| 1709 |
+
},
|
| 1710 |
+
"secs": 22.472158670425415,
|
| 1711 |
+
"rl_secs_used": 1550.43736743927
|
| 1712 |
+
},
|
| 1713 |
+
{
|
| 1714 |
+
"round": 53,
|
| 1715 |
+
"level": 1,
|
| 1716 |
+
"k": 8,
|
| 1717 |
+
"n_targets": 1500,
|
| 1718 |
+
"sample_acc": 0.6666666666666666,
|
| 1719 |
+
"new_proofs": 1089,
|
| 1720 |
+
"targets_solved": 1313,
|
| 1721 |
+
"lstar": 11,
|
| 1722 |
+
"ge": {
|
| 1723 |
+
"7": 1313,
|
| 1724 |
+
"8": 718,
|
| 1725 |
+
"9": 406,
|
| 1726 |
+
"10": 121,
|
| 1727 |
+
"11": 25,
|
| 1728 |
+
"12": 4,
|
| 1729 |
+
"13": 0,
|
| 1730 |
+
"14": 0
|
| 1731 |
+
},
|
| 1732 |
+
"mix_rl_records": 4165,
|
| 1733 |
+
"sampling": {
|
| 1734 |
+
"samples": 12000,
|
| 1735 |
+
"secs": 63.54386305809021,
|
| 1736 |
+
"gate_secs": 30.18801736831665,
|
| 1737 |
+
"samples_per_sec": 188.84593133769505
|
| 1738 |
+
},
|
| 1739 |
+
"secs": 105.78336071968079,
|
| 1740 |
+
"rl_secs_used": 1626.0329568386078
|
| 1741 |
+
},
|
| 1742 |
+
{
|
| 1743 |
+
"round": 54,
|
| 1744 |
+
"level": 2,
|
| 1745 |
+
"k": 24,
|
| 1746 |
+
"n_targets": 76,
|
| 1747 |
+
"sample_acc": 0.08662280701754387,
|
| 1748 |
+
"new_proofs": 49,
|
| 1749 |
+
"targets_solved": 1313,
|
| 1750 |
+
"lstar": 11,
|
| 1751 |
+
"ge": {
|
| 1752 |
+
"7": 1313,
|
| 1753 |
+
"8": 718,
|
| 1754 |
+
"9": 406,
|
| 1755 |
+
"10": 121,
|
| 1756 |
+
"11": 25,
|
| 1757 |
+
"12": 4,
|
| 1758 |
+
"13": 0,
|
| 1759 |
+
"14": 0
|
| 1760 |
+
},
|
| 1761 |
+
"mix_rl_records": 79,
|
| 1762 |
+
"sampling": {
|
| 1763 |
+
"samples": 1824,
|
| 1764 |
+
"secs": 12.209544658660889,
|
| 1765 |
+
"gate_secs": 4.265426874160767,
|
| 1766 |
+
"samples_per_sec": 149.39132056052054
|
| 1767 |
+
},
|
| 1768 |
+
"secs": 12.280830383300781,
|
| 1769 |
+
"rl_secs_used": 1634.0481152534485
|
| 1770 |
+
},
|
| 1771 |
+
{
|
| 1772 |
+
"round": 55,
|
| 1773 |
+
"level": 4,
|
| 1774 |
+
"k": 24,
|
| 1775 |
+
"n_targets": 75,
|
| 1776 |
+
"sample_acc": 0.09277777777777778,
|
| 1777 |
+
"new_proofs": 46,
|
| 1778 |
+
"targets_solved": 1314,
|
| 1779 |
+
"lstar": 11,
|
| 1780 |
+
"ge": {
|
| 1781 |
+
"7": 1314,
|
| 1782 |
+
"8": 718,
|
| 1783 |
+
"9": 406,
|
| 1784 |
+
"10": 121,
|
| 1785 |
+
"11": 25,
|
| 1786 |
+
"12": 4,
|
| 1787 |
+
"13": 0,
|
| 1788 |
+
"14": 0
|
| 1789 |
+
},
|
| 1790 |
+
"mix_rl_records": 70,
|
| 1791 |
+
"sampling": {
|
| 1792 |
+
"samples": 1800,
|
| 1793 |
+
"secs": 16.715649604797363,
|
| 1794 |
+
"gate_secs": 4.045336723327637,
|
| 1795 |
+
"samples_per_sec": 107.68352068611219
|
| 1796 |
+
},
|
| 1797 |
+
"secs": 16.780733346939087,
|
| 1798 |
+
"rl_secs_used": 1646.7835109233856
|
| 1799 |
+
},
|
| 1800 |
+
{
|
| 1801 |
+
"round": 56,
|
| 1802 |
+
"level": 8,
|
| 1803 |
+
"k": 24,
|
| 1804 |
+
"n_targets": 75,
|
| 1805 |
+
"sample_acc": 0.07888888888888888,
|
| 1806 |
+
"new_proofs": 37,
|
| 1807 |
+
"targets_solved": 1314,
|
| 1808 |
+
"lstar": 11,
|
| 1809 |
+
"ge": {
|
| 1810 |
+
"7": 1314,
|
| 1811 |
+
"8": 718,
|
| 1812 |
+
"9": 406,
|
| 1813 |
+
"10": 121,
|
| 1814 |
+
"11": 25,
|
| 1815 |
+
"12": 4,
|
| 1816 |
+
"13": 0,
|
| 1817 |
+
"14": 0
|
| 1818 |
+
},
|
| 1819 |
+
"mix_rl_records": 72,
|
| 1820 |
+
"sampling": {
|
| 1821 |
+
"samples": 1800,
|
| 1822 |
+
"secs": 25.34758234024048,
|
| 1823 |
+
"gate_secs": 4.270497560501099,
|
| 1824 |
+
"samples_per_sec": 71.01268972474804
|
| 1825 |
+
},
|
| 1826 |
+
"secs": 25.412888050079346,
|
| 1827 |
+
"rl_secs_used": 1667.925907611847
|
| 1828 |
+
}
|
| 1829 |
+
],
|
| 1830 |
+
"secs": 2416.636120080948,
|
| 1831 |
+
"rl_secs_used": 1709.9186789989471,
|
| 1832 |
+
"gate_secs": 706.6594099998474,
|
| 1833 |
+
"targets": 1500,
|
| 1834 |
+
"stats": {
|
| 1835 |
+
"rounds": 14,
|
| 1836 |
+
"first_loops": 1,
|
| 1837 |
+
"k_first": 8,
|
| 1838 |
+
"retry_loops": [
|
| 1839 |
+
2,
|
| 1840 |
+
4,
|
| 1841 |
+
8
|
| 1842 |
+
],
|
| 1843 |
+
"k_retry": 24,
|
| 1844 |
+
"retries": [
|
| 1845 |
+
{
|
| 1846 |
+
"round": 1,
|
| 1847 |
+
"loops": 2,
|
| 1848 |
+
"targets": 365,
|
| 1849 |
+
"cracked": 92
|
| 1850 |
+
},
|
| 1851 |
+
{
|
| 1852 |
+
"round": 1,
|
| 1853 |
+
"loops": 4,
|
| 1854 |
+
"targets": 365,
|
| 1855 |
+
"cracked": 109
|
| 1856 |
+
},
|
| 1857 |
+
{
|
| 1858 |
+
"round": 1,
|
| 1859 |
+
"loops": 8,
|
| 1860 |
+
"targets": 364,
|
| 1861 |
+
"cracked": 116
|
| 1862 |
+
},
|
| 1863 |
+
{
|
| 1864 |
+
"round": 2,
|
| 1865 |
+
"loops": 2,
|
| 1866 |
+
"targets": 227,
|
| 1867 |
+
"cracked": 36
|
| 1868 |
+
},
|
| 1869 |
+
{
|
| 1870 |
+
"round": 2,
|
| 1871 |
+
"loops": 4,
|
| 1872 |
+
"targets": 227,
|
| 1873 |
+
"cracked": 63
|
| 1874 |
+
},
|
| 1875 |
+
{
|
| 1876 |
+
"round": 2,
|
| 1877 |
+
"loops": 8,
|
| 1878 |
+
"targets": 226,
|
| 1879 |
+
"cracked": 67
|
| 1880 |
+
},
|
| 1881 |
+
{
|
| 1882 |
+
"round": 3,
|
| 1883 |
+
"loops": 2,
|
| 1884 |
+
"targets": 181,
|
| 1885 |
+
"cracked": 32
|
| 1886 |
+
},
|
| 1887 |
+
{
|
| 1888 |
+
"round": 3,
|
| 1889 |
+
"loops": 4,
|
| 1890 |
+
"targets": 181,
|
| 1891 |
+
"cracked": 44
|
| 1892 |
+
},
|
| 1893 |
+
{
|
| 1894 |
+
"round": 3,
|
| 1895 |
+
"loops": 8,
|
| 1896 |
+
"targets": 180,
|
| 1897 |
+
"cracked": 49
|
| 1898 |
+
},
|
| 1899 |
+
{
|
| 1900 |
+
"round": 4,
|
| 1901 |
+
"loops": 2,
|
| 1902 |
+
"targets": 147,
|
| 1903 |
+
"cracked": 30
|
| 1904 |
+
},
|
| 1905 |
+
{
|
| 1906 |
+
"round": 4,
|
| 1907 |
+
"loops": 4,
|
| 1908 |
+
"targets": 147,
|
| 1909 |
+
"cracked": 31
|
| 1910 |
+
},
|
| 1911 |
+
{
|
| 1912 |
+
"round": 4,
|
| 1913 |
+
"loops": 8,
|
| 1914 |
+
"targets": 146,
|
| 1915 |
+
"cracked": 42
|
| 1916 |
+
},
|
| 1917 |
+
{
|
| 1918 |
+
"round": 5,
|
| 1919 |
+
"loops": 2,
|
| 1920 |
+
"targets": 131,
|
| 1921 |
+
"cracked": 28
|
| 1922 |
+
},
|
| 1923 |
+
{
|
| 1924 |
+
"round": 5,
|
| 1925 |
+
"loops": 4,
|
| 1926 |
+
"targets": 130,
|
| 1927 |
+
"cracked": 37
|
| 1928 |
+
},
|
| 1929 |
+
{
|
| 1930 |
+
"round": 5,
|
| 1931 |
+
"loops": 8,
|
| 1932 |
+
"targets": 130,
|
| 1933 |
+
"cracked": 33
|
| 1934 |
+
},
|
| 1935 |
+
{
|
| 1936 |
+
"round": 6,
|
| 1937 |
+
"loops": 2,
|
| 1938 |
+
"targets": 110,
|
| 1939 |
+
"cracked": 27
|
| 1940 |
+
},
|
| 1941 |
+
{
|
| 1942 |
+
"round": 6,
|
| 1943 |
+
"loops": 4,
|
| 1944 |
+
"targets": 110,
|
| 1945 |
+
"cracked": 28
|
| 1946 |
+
},
|
| 1947 |
+
{
|
| 1948 |
+
"round": 6,
|
| 1949 |
+
"loops": 8,
|
| 1950 |
+
"targets": 109,
|
| 1951 |
+
"cracked": 25
|
| 1952 |
+
},
|
| 1953 |
+
{
|
| 1954 |
+
"round": 7,
|
| 1955 |
+
"loops": 2,
|
| 1956 |
+
"targets": 102,
|
| 1957 |
+
"cracked": 27
|
| 1958 |
+
},
|
| 1959 |
+
{
|
| 1960 |
+
"round": 7,
|
| 1961 |
+
"loops": 4,
|
| 1962 |
+
"targets": 102,
|
| 1963 |
+
"cracked": 23
|
| 1964 |
+
},
|
| 1965 |
+
{
|
| 1966 |
+
"round": 7,
|
| 1967 |
+
"loops": 8,
|
| 1968 |
+
"targets": 101,
|
| 1969 |
+
"cracked": 26
|
| 1970 |
+
},
|
| 1971 |
+
{
|
| 1972 |
+
"round": 8,
|
| 1973 |
+
"loops": 2,
|
| 1974 |
+
"targets": 95,
|
| 1975 |
+
"cracked": 21
|
| 1976 |
+
},
|
| 1977 |
+
{
|
| 1978 |
+
"round": 8,
|
| 1979 |
+
"loops": 4,
|
| 1980 |
+
"targets": 95,
|
| 1981 |
+
"cracked": 25
|
| 1982 |
+
},
|
| 1983 |
+
{
|
| 1984 |
+
"round": 8,
|
| 1985 |
+
"loops": 8,
|
| 1986 |
+
"targets": 95,
|
| 1987 |
+
"cracked": 17
|
| 1988 |
+
},
|
| 1989 |
+
{
|
| 1990 |
+
"round": 9,
|
| 1991 |
+
"loops": 2,
|
| 1992 |
+
"targets": 86,
|
| 1993 |
+
"cracked": 13
|
| 1994 |
+
},
|
| 1995 |
+
{
|
| 1996 |
+
"round": 9,
|
| 1997 |
+
"loops": 4,
|
| 1998 |
+
"targets": 86,
|
| 1999 |
+
"cracked": 14
|
| 2000 |
+
},
|
| 2001 |
+
{
|
| 2002 |
+
"round": 9,
|
| 2003 |
+
"loops": 8,
|
| 2004 |
+
"targets": 85,
|
| 2005 |
+
"cracked": 18
|
| 2006 |
+
},
|
| 2007 |
+
{
|
| 2008 |
+
"round": 10,
|
| 2009 |
+
"loops": 2,
|
| 2010 |
+
"targets": 84,
|
| 2011 |
+
"cracked": 19
|
| 2012 |
+
},
|
| 2013 |
+
{
|
| 2014 |
+
"round": 10,
|
| 2015 |
+
"loops": 4,
|
| 2016 |
+
"targets": 84,
|
| 2017 |
+
"cracked": 12
|
| 2018 |
+
},
|
| 2019 |
+
{
|
| 2020 |
+
"round": 10,
|
| 2021 |
+
"loops": 8,
|
| 2022 |
+
"targets": 83,
|
| 2023 |
+
"cracked": 12
|
| 2024 |
+
},
|
| 2025 |
+
{
|
| 2026 |
+
"round": 11,
|
| 2027 |
+
"loops": 2,
|
| 2028 |
+
"targets": 78,
|
| 2029 |
+
"cracked": 11
|
| 2030 |
+
},
|
| 2031 |
+
{
|
| 2032 |
+
"round": 11,
|
| 2033 |
+
"loops": 4,
|
| 2034 |
+
"targets": 78,
|
| 2035 |
+
"cracked": 11
|
| 2036 |
+
},
|
| 2037 |
+
{
|
| 2038 |
+
"round": 11,
|
| 2039 |
+
"loops": 8,
|
| 2040 |
+
"targets": 78,
|
| 2041 |
+
"cracked": 13
|
| 2042 |
+
},
|
| 2043 |
+
{
|
| 2044 |
+
"round": 12,
|
| 2045 |
+
"loops": 2,
|
| 2046 |
+
"targets": 79,
|
| 2047 |
+
"cracked": 10
|
| 2048 |
+
},
|
| 2049 |
+
{
|
| 2050 |
+
"round": 12,
|
| 2051 |
+
"loops": 4,
|
| 2052 |
+
"targets": 78,
|
| 2053 |
+
"cracked": 12
|
| 2054 |
+
},
|
| 2055 |
+
{
|
| 2056 |
+
"round": 12,
|
| 2057 |
+
"loops": 8,
|
| 2058 |
+
"targets": 78,
|
| 2059 |
+
"cracked": 17
|
| 2060 |
+
},
|
| 2061 |
+
{
|
| 2062 |
+
"round": 13,
|
| 2063 |
+
"loops": 2,
|
| 2064 |
+
"targets": 75,
|
| 2065 |
+
"cracked": 9
|
| 2066 |
+
},
|
| 2067 |
+
{
|
| 2068 |
+
"round": 13,
|
| 2069 |
+
"loops": 4,
|
| 2070 |
+
"targets": 75,
|
| 2071 |
+
"cracked": 15
|
| 2072 |
+
},
|
| 2073 |
+
{
|
| 2074 |
+
"round": 13,
|
| 2075 |
+
"loops": 8,
|
| 2076 |
+
"targets": 74,
|
| 2077 |
+
"cracked": 9
|
| 2078 |
+
},
|
| 2079 |
+
{
|
| 2080 |
+
"round": 14,
|
| 2081 |
+
"loops": 2,
|
| 2082 |
+
"targets": 76,
|
| 2083 |
+
"cracked": 12
|
| 2084 |
+
},
|
| 2085 |
+
{
|
| 2086 |
+
"round": 14,
|
| 2087 |
+
"loops": 4,
|
| 2088 |
+
"targets": 75,
|
| 2089 |
+
"cracked": 13
|
| 2090 |
+
},
|
| 2091 |
+
{
|
| 2092 |
+
"round": 14,
|
| 2093 |
+
"loops": 8,
|
| 2094 |
+
"targets": 75,
|
| 2095 |
+
"cracked": 10
|
| 2096 |
+
}
|
| 2097 |
+
],
|
| 2098 |
+
"ft_records": 38400,
|
| 2099 |
+
"ft_steps": [
|
| 2100 |
+
313,
|
| 2101 |
+
329,
|
| 2102 |
+
341,
|
| 2103 |
+
349,
|
| 2104 |
+
355,
|
| 2105 |
+
354,
|
| 2106 |
+
373,
|
| 2107 |
+
373,
|
| 2108 |
+
380,
|
| 2109 |
+
384,
|
| 2110 |
+
391,
|
| 2111 |
+
390,
|
| 2112 |
+
399,
|
| 2113 |
+
399
|
| 2114 |
+
],
|
| 2115 |
+
"ft_optimizer": "muon (polar express) on block matrices and head + adamw on the rest",
|
| 2116 |
+
"ft_lr": 0.0003,
|
| 2117 |
+
"ft_muon_lr": 0.003
|
| 2118 |
+
}
|
| 2119 |
+
},
|
| 2120 |
+
"eval": {
|
| 2121 |
+
"level": 8,
|
| 2122 |
+
"attempts_per_theorem": 64,
|
| 2123 |
+
"solved": 815,
|
| 2124 |
+
"n": 1108,
|
| 2125 |
+
"solved_ge7": 815,
|
| 2126 |
+
"lstar": 12,
|
| 2127 |
+
"ge": {
|
| 2128 |
+
"7": 815,
|
| 2129 |
+
"8": 683,
|
| 2130 |
+
"9": 547,
|
| 2131 |
+
"10": 147,
|
| 2132 |
+
"11": 34,
|
| 2133 |
+
"12": 8,
|
| 2134 |
+
"13": 1,
|
| 2135 |
+
"14": 0
|
| 2136 |
+
},
|
| 2137 |
+
"by_bin": {
|
| 2138 |
+
"7": {
|
| 2139 |
+
"solved": 132,
|
| 2140 |
+
"n": 150,
|
| 2141 |
+
"rate": 0.88,
|
| 2142 |
+
"ci": [
|
| 2143 |
+
0.8182903386460476,
|
| 2144 |
+
0.9227315956032063
|
| 2145 |
+
]
|
| 2146 |
+
},
|
| 2147 |
+
"8": {
|
| 2148 |
+
"solved": 136,
|
| 2149 |
+
"n": 152,
|
| 2150 |
+
"rate": 0.8947368421052632,
|
| 2151 |
+
"ci": [
|
| 2152 |
+
0.8358499003389465,
|
| 2153 |
+
0.9341627278681562
|
| 2154 |
+
]
|
| 2155 |
+
},
|
| 2156 |
+
"9": {
|
| 2157 |
+
"solved": 400,
|
| 2158 |
+
"n": 477,
|
| 2159 |
+
"rate": 0.8385744234800838,
|
| 2160 |
+
"ci": [
|
| 2161 |
+
0.8028723089318024,
|
| 2162 |
+
0.8688665755573932
|
| 2163 |
+
]
|
| 2164 |
+
},
|
| 2165 |
+
"10": {
|
| 2166 |
+
"solved": 113,
|
| 2167 |
+
"n": 219,
|
| 2168 |
+
"rate": 0.5159817351598174,
|
| 2169 |
+
"ci": [
|
| 2170 |
+
0.4500902155545144,
|
| 2171 |
+
0.5813222316725744
|
| 2172 |
+
]
|
| 2173 |
+
},
|
| 2174 |
+
"11": {
|
| 2175 |
+
"solved": 26,
|
| 2176 |
+
"n": 49,
|
| 2177 |
+
"rate": 0.5306122448979592,
|
| 2178 |
+
"ci": [
|
| 2179 |
+
0.39380627019919845,
|
| 2180 |
+
0.6629671810210522
|
| 2181 |
+
]
|
| 2182 |
+
},
|
| 2183 |
+
"12": {
|
| 2184 |
+
"solved": 7,
|
| 2185 |
+
"n": 54,
|
| 2186 |
+
"rate": 0.12962962962962962,
|
| 2187 |
+
"ci": [
|
| 2188 |
+
0.06423639229487982,
|
| 2189 |
+
0.2442198053206771
|
| 2190 |
+
]
|
| 2191 |
+
},
|
| 2192 |
+
"13": {
|
| 2193 |
+
"solved": 1,
|
| 2194 |
+
"n": 5,
|
| 2195 |
+
"rate": 0.2,
|
| 2196 |
+
"ci": [
|
| 2197 |
+
0.03622316096978745,
|
| 2198 |
+
0.6244717358814613
|
| 2199 |
+
]
|
| 2200 |
+
},
|
| 2201 |
+
"14": {
|
| 2202 |
+
"solved": 0,
|
| 2203 |
+
"n": 2,
|
| 2204 |
+
"rate": 0.0,
|
| 2205 |
+
"ci": [
|
| 2206 |
+
0.0,
|
| 2207 |
+
0.6576280471103807
|
| 2208 |
+
]
|
| 2209 |
+
}
|
| 2210 |
+
},
|
| 2211 |
+
"sample_acc": 0.4808212996389892,
|
| 2212 |
+
"max_pruned": 16,
|
| 2213 |
+
"sampling": {
|
| 2214 |
+
"samples": 70912,
|
| 2215 |
+
"secs": 816.1614382266998,
|
| 2216 |
+
"gate_secs": 132.59109663963318,
|
| 2217 |
+
"samples_per_sec": 86.88477141737152
|
| 2218 |
+
},
|
| 2219 |
+
"heldout_greedy": 0.9855,
|
| 2220 |
+
"diag": {
|
| 2221 |
+
"level": 2,
|
| 2222 |
+
"attempts_per_theorem": 64,
|
| 2223 |
+
"solved": 795,
|
| 2224 |
+
"n": 1108,
|
| 2225 |
+
"solved_ge7": 795,
|
| 2226 |
+
"lstar": 12,
|
| 2227 |
+
"ge": {
|
| 2228 |
+
"7": 795,
|
| 2229 |
+
"8": 673,
|
| 2230 |
+
"9": 538,
|
| 2231 |
+
"10": 147,
|
| 2232 |
+
"11": 35,
|
| 2233 |
+
"12": 9,
|
| 2234 |
+
"13": 2,
|
| 2235 |
+
"14": 1
|
| 2236 |
+
},
|
| 2237 |
+
"by_bin": {
|
| 2238 |
+
"7": {
|
| 2239 |
+
"solved": 122,
|
| 2240 |
+
"n": 150,
|
| 2241 |
+
"rate": 0.8133333333333334,
|
| 2242 |
+
"ci": [
|
| 2243 |
+
0.7434413916324496,
|
| 2244 |
+
0.8675766945029
|
| 2245 |
+
]
|
| 2246 |
+
},
|
| 2247 |
+
"8": {
|
| 2248 |
+
"solved": 135,
|
| 2249 |
+
"n": 152,
|
| 2250 |
+
"rate": 0.8881578947368421,
|
| 2251 |
+
"ci": [
|
| 2252 |
+
0.8281892492606145,
|
| 2253 |
+
0.9289898351430365
|
| 2254 |
+
]
|
| 2255 |
+
},
|
| 2256 |
+
"9": {
|
| 2257 |
+
"solved": 391,
|
| 2258 |
+
"n": 477,
|
| 2259 |
+
"rate": 0.8197064989517819,
|
| 2260 |
+
"ci": [
|
| 2261 |
+
0.7826957857606539,
|
| 2262 |
+
0.8516087336486483
|
| 2263 |
+
]
|
| 2264 |
+
},
|
| 2265 |
+
"10": {
|
| 2266 |
+
"solved": 112,
|
| 2267 |
+
"n": 219,
|
| 2268 |
+
"rate": 0.5114155251141552,
|
| 2269 |
+
"ci": [
|
| 2270 |
+
0.4455865745445954,
|
| 2271 |
+
0.5768508877604678
|
| 2272 |
+
]
|
| 2273 |
+
},
|
| 2274 |
+
"11": {
|
| 2275 |
+
"solved": 26,
|
| 2276 |
+
"n": 49,
|
| 2277 |
+
"rate": 0.5306122448979592,
|
| 2278 |
+
"ci": [
|
| 2279 |
+
0.39380627019919845,
|
| 2280 |
+
0.6629671810210522
|
| 2281 |
+
]
|
| 2282 |
+
},
|
| 2283 |
+
"12": {
|
| 2284 |
+
"solved": 7,
|
| 2285 |
+
"n": 54,
|
| 2286 |
+
"rate": 0.12962962962962962,
|
| 2287 |
+
"ci": [
|
| 2288 |
+
0.06423639229487982,
|
| 2289 |
+
0.2442198053206771
|
| 2290 |
+
]
|
| 2291 |
+
},
|
| 2292 |
+
"13": {
|
| 2293 |
+
"solved": 1,
|
| 2294 |
+
"n": 5,
|
| 2295 |
+
"rate": 0.2,
|
| 2296 |
+
"ci": [
|
| 2297 |
+
0.03622316096978745,
|
| 2298 |
+
0.6244717358814613
|
| 2299 |
+
]
|
| 2300 |
+
},
|
| 2301 |
+
"14": {
|
| 2302 |
+
"solved": 1,
|
| 2303 |
+
"n": 2,
|
| 2304 |
+
"rate": 0.5,
|
| 2305 |
+
"ci": [
|
| 2306 |
+
0.09452865480086614,
|
| 2307 |
+
0.9054713451991339
|
| 2308 |
+
]
|
| 2309 |
+
}
|
| 2310 |
+
},
|
| 2311 |
+
"sample_acc": 0.44345103790613716,
|
| 2312 |
+
"max_pruned": 17,
|
| 2313 |
+
"sampling": {
|
| 2314 |
+
"samples": 70912,
|
| 2315 |
+
"secs": 380.8417167663574,
|
| 2316 |
+
"gate_secs": 129.8440375328064,
|
| 2317 |
+
"samples_per_sec": 186.19808933248717
|
| 2318 |
+
}
|
| 2319 |
+
}
|
| 2320 |
+
}
|
| 2321 |
+
}
|