Delta-Vector commited on
Commit
bddbc01
·
verified ·
1 Parent(s): ee52a36

Upload provenance/train.toml with huggingface_hub

Browse files
Files changed (1) hide show
  1. provenance/train.toml +79 -0
provenance/train.toml ADDED
@@ -0,0 +1,79 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # Trinity-Mini SFT: Nemotron-3-Ultra-SFT-Blend + Dans-Benchmaxx + Dans-Benchmaxx-COT
2
+ # Hparams from sweep winner (muon-8e6, fgn20xry) in trinity-mini-sft-sweeps.
3
+ # seq_len doubled 32768 -> 65536 with batch_size halved 16 -> 8: identical
4
+ # 524,288 tokens/step, far less packing truncation (data p90 ~43k tokens).
5
+ # max_steps = 1 epoch, computed from prep_stats.json packing simulation.
6
+
7
+ output_dir = "/home/ubuntu/prime/runs/prod-v3"
8
+ max_steps = 8136 # 2 epochs (user directive): 2 x 4,068 steps/epoch (32,547 bins / 8, prep_stats.json)
9
+ loss_impl = "liger_fused"
10
+ dist_timeout_seconds = 1200
11
+
12
+ [deployment]
13
+ num_gpus = 8
14
+
15
+ [model]
16
+ name = "arcee-ai/Trinity-Mini"
17
+ impl = "custom"
18
+ attn = "flash_attention_3"
19
+ seq_len = 65536
20
+
21
+ [model.compile]
22
+
23
+ [model.ac]
24
+ mode = "full"
25
+ freq = 1
26
+
27
+ [tokenizer]
28
+ chat_template = "/home/ubuntu/prime/prime-rl/configs/trinity_sft/trinity_sft.jinja"
29
+
30
+ [data]
31
+ type = "sft"
32
+ name = "/home/ubuntu/prime/data/train_ds"
33
+ batch_size = 8
34
+ micro_batch_size = 1
35
+ seq_len = 65536
36
+ pack_function = "cat"
37
+
38
+ [data.loss_mask]
39
+ # Constant-reasoner: non-thinking samples' leading "</think>\n\n" is treated as
40
+ # scaffolding (masked) so the model never learns to skip thinking unprompted.
41
+ immediate_think_close = false
42
+
43
+ [val]
44
+ interval = 250
45
+ eval_on_start = true
46
+
47
+ [val.data]
48
+ type = "sft"
49
+ name = "/home/ubuntu/prime/data/val_ds"
50
+ batch_size = 8
51
+ micro_batch_size = 1
52
+ seq_len = 65536
53
+
54
+ [val.data.loss_mask]
55
+ immediate_think_close = false
56
+
57
+ [optim]
58
+ type = "muon"
59
+ lr = 8e-6
60
+ mu = 0.95
61
+ betas1 = 0.9
62
+ betas2 = 0.95
63
+ weight_decay = 0.001
64
+ max_norm = 1.0
65
+
66
+ [scheduler]
67
+ type = "cosine"
68
+ warmup_steps = 100
69
+ min_lr = 1e-7
70
+
71
+ [ckpt]
72
+ interval = 250
73
+ keep_last = 1
74
+
75
+ [wandb]
76
+ project = "trinity-mini-sft-sweeps"
77
+ group = "prod-v3"
78
+ name = "prod-nemotron-ultra-benchmaxx-8e6-64k-maskclose"
79
+ tags = ["trinity-mini", "afmoe", "muon", "lr8e-6", "prod", "nemotron-ultra", "benchmaxx", "seq64k", "autonomous", "mask-think-close"]