ekdeepslubana commited on
Commit
fd6d6f5
·
verified ·
1 Parent(s): c4f81c8

Upload folder using huggingface_hub

Browse files
batchtopk_100M/conf.yaml ADDED
@@ -0,0 +1,43 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ data:
2
+ batch_size: 100
3
+ cache_dir: ../../activations/precomputed_activations/precomputed_activations_gemma2/
4
+ context_length: 500
5
+ dtype: bfloat16
6
+ epochs: 1
7
+ hf_name: monology/pile-uncopyrighted
8
+ num_total_steps: 200000
9
+ num_workers: 2
10
+ deploy: true
11
+ device_id: cuda:0
12
+ eval:
13
+ save_tables: false
14
+ llm:
15
+ dimin: 2304
16
+ model_hf_name: google/gemma-2-2b
17
+ tokenizer_hf_name: google/gemma-2-2b
18
+ log:
19
+ log_interval: 10
20
+ save_interval: 20000
21
+ save_multiple: true
22
+ wandb_project_name: TemporalSAE_test
23
+ optimizer:
24
+ beta1: 0.9
25
+ beta2: 0.95
26
+ decay_lr: true
27
+ grad_clip: 1.0
28
+ learning_rate: 0.001
29
+ min_lr: 0.0009
30
+ warmup_iters: 200
31
+ weight_decay: 0.0001
32
+ sae:
33
+ block_id: 1
34
+ branchingFactor: 10
35
+ encoder_reg: true
36
+ exp_factor: 4
37
+ gamma_reg: 5
38
+ kval_topk: 192
39
+ mp_kval: 192
40
+ sae_type: batchtopk
41
+ scaling_factor: 0.00666666667
42
+ seed: 42
43
+ tag: scratch
batchtopk_100M/latest_ckpt.pt ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:cfeb18a9bad02b44c7925035d87a9b6148265abdb870e6c4d01c5c11daee56ff
3
+ size 509683298