Shasidharyadavr commited on
Commit
cddc933
·
verified ·
1 Parent(s): 86a4c24

GRPO 3B v24 (temp=0.7, LR=2e-05, r=32, model=Qwen2.5-Math-7B-Instruct): 0.573->0.534 delta=-0.039

Browse files
Files changed (1) hide show
  1. training_results_grpo_3b_v24.json +89 -0
training_results_grpo_3b_v24.json ADDED
@@ -0,0 +1,89 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "env_url": "https://shasidharyadavr-netweaver-sre.hf.space",
3
+ "model_name": "Qwen/Qwen2.5-Math-7B-Instruct",
4
+ "training_method": "GRPO v24: 4-bit + LoRA r=32 (q_proj+k_proj+v_proj+o_proj+gate_proj+up_proj+down_proj), LR=2e-05, temp_train=temp_eval=0.7",
5
+ "max_train_steps": 25,
6
+ "num_generations": 4,
7
+ "learning_rate": 2e-05,
8
+ "lora_rank": 32,
9
+ "lora_target_modules": [
10
+ "q_proj",
11
+ "k_proj",
12
+ "v_proj",
13
+ "o_proj",
14
+ "gate_proj",
15
+ "up_proj",
16
+ "down_proj"
17
+ ],
18
+ "training_temperature": 0.7,
19
+ "eval_temperature": 0.7,
20
+ "trainable_params": 80740352,
21
+ "total_params": 4433712640,
22
+ "training_rewards": [
23
+ 0.9,
24
+ 0.667,
25
+ 0.5,
26
+ 0.9,
27
+ 0.3,
28
+ 0.9,
29
+ 0.9,
30
+ 0.9,
31
+ 0.5,
32
+ 0.9,
33
+ 0.9,
34
+ 0.3,
35
+ 0.3,
36
+ 0.5,
37
+ 0.5,
38
+ 0.5,
39
+ 0.3,
40
+ 0.3,
41
+ 0.5,
42
+ 0.3,
43
+ 0.5,
44
+ 0.3,
45
+ 0.3,
46
+ 0.9,
47
+ 0.3
48
+ ],
49
+ "before_rewards": [
50
+ 0.3,
51
+ 0.3,
52
+ 0.9,
53
+ 0.4,
54
+ 0.667,
55
+ 0.667,
56
+ 0.5,
57
+ 0.9,
58
+ 0.6,
59
+ 0.5
60
+ ],
61
+ "after_rewards": [
62
+ 0.3,
63
+ 0.88,
64
+ 0.3,
65
+ 0.9,
66
+ 0.3,
67
+ 0.86,
68
+ 0.5,
69
+ 0.5,
70
+ 0.5,
71
+ 0.3
72
+ ],
73
+ "difficulty_breakdown": {
74
+ "before": {
75
+ "easy": 0.45,
76
+ "medium": 0.7835000000000001,
77
+ "hard": 0.425
78
+ },
79
+ "after": {
80
+ "easy": 0.545,
81
+ "medium": 0.6866666666666666,
82
+ "hard": 0.3666666666666667
83
+ }
84
+ },
85
+ "timestamp": "2026-04-26 12:14:18",
86
+ "source": "grpo_3b_real_run_v24",
87
+ "job_flavor": "a10g-large",
88
+ "notes": "GRPO v24 on Qwen/Qwen2.5-Math-7B-Instruct (4-bit + LoRA r=32, targets=['q_proj', 'k_proj', 'v_proj', 'o_proj', 'gate_proj', 'up_proj', 'down_proj']) vs the live NetWeaver SRE env. Train temp = eval temp = 0.7, LR = 2e-05. Heuristic baseline preserved (separate filenames)."
89
+ }