GRRNMAKER commited on
Commit
ed62455
·
verified ·
1 Parent(s): 0323459

Upload reasoning_registry.py with huggingface_hub

Browse files
Files changed (1) hide show
  1. reasoning_registry.py +67 -0
reasoning_registry.py ADDED
@@ -0,0 +1,67 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ """
2
+ Cozet Phase 2: NVIDIA Reasoning Collection (Code + Math + General)
3
+ ====================================================================
4
+
5
+ CODE REASONING DATASETS:
6
+ nvidia/OpenCodeReasoning 753k samples (v1)
7
+ nvidia/OpenCodeReasoning-2 2.16M samples (v2, self-critique)
8
+ Paper v1: arxiv:2504.01943
9
+ Paper v2: arxiv:2507.09075
10
+
11
+ MATH REASONING DATASET:
12
+ nvidia/OpenMathReasoning 5.68M samples
13
+ Paper: arxiv:2504.16891 (AIMO-2 winner)
14
+
15
+ REFERENCE MODELS (for distillation):
16
+ Code reasoning:
17
+ nvidia/OpenCodeReasoning-Nemotron-1.1-7B 8B (latest v1.1)
18
+ nvidia/OpenCodeReasoning-Nemotron-1.1-14B 15B
19
+ nvidia/OpenCodeReasoning-Nemotron-1.1-32B 33B
20
+ nvidia/OpenCodeReasoning-Nemotron-32B-IOI 33B (IOI competition)
21
+ Math reasoning:
22
+ nvidia/AceMath-RL-Nemotron-7B RL math
23
+ nvidia/OpenMath-Nemotron-7B 8B
24
+ nvidia/OpenMath-Nemotron-32B 33B
25
+ General reasoning:
26
+ nvidia/OpenReasoning-Nemotron-7B 8B
27
+ nvidia/OpenReasoning-Nemotron-14B 15B
28
+ nvidia/OpenReasoning-Nemotron-32B 33B (67k downloads)
29
+
30
+ TOTAL REASONING TRAINING DATA:
31
+ Code: 2.16M + 753k = 2.91M samples
32
+ Math: 5.68M samples
33
+ Total: 8.59M reasoning traces
34
+
35
+ COZET FULL PHASE 2 DATA STACK (COMPLETE):
36
+ -------------------------------------------------------
37
+ Category Dataset Samples
38
+ -------------------------------------------------------
39
+ Language FineWeb-Edu 10B tokens
40
+ Language OCR-Synthetic-Multilingual 11M images
41
+ Language Nigerian (Yoruba/Hausa/Igbo) TBD
42
+ -------------------------------------------------------
43
+ Coding Pretraining-Code-v2 836M
44
+ Coding CC-Code-v1 216M
45
+ Coding SFT-SWE-v3 238k
46
+ Coding Cascade-SFT-SWE 141k
47
+ Coding Cascade-RL-SWE 110k
48
+ Coding SWE-Zero/Hero/Open traces varies
49
+ Coding SFT-CUDA-v1 2.28k
50
+ -------------------------------------------------------
51
+ Reasoning-Code OpenCodeReasoning-2 2.16M
52
+ Reasoning-Code OpenCodeReasoning 753k
53
+ Reasoning-Math OpenMathReasoning 5.68M
54
+ Reasoning-AGI ARC-AGI 122k
55
+ -------------------------------------------------------
56
+ Agentic Conversational-Tool-Use 97k
57
+ Agentic Terminal-Corpus 366k
58
+ Agentic Function-Calling 9.62k
59
+ Agentic SWE-Pivot 33.8k
60
+ Agentic Calendar/Workplace 5.8k
61
+ -------------------------------------------------------
62
+ Safety Prompt-Injection 1.27k
63
+ -------------------------------------------------------
64
+ Augment SSoT (10% entropy prefix) applied
65
+ -------------------------------------------------------
66
+ TOTAL: ~1B+ pretraining + ~9M+ reasoning + ~850k agentic
67
+ """