Hugging Face's logo

Zishan-Shao
/
decodeshare

interpretability
mechanistic-interpretability
activation-steering
llm
kv-cache
icml-2026
Model card Files
xet
 Community
Zishan-Shao commited on
Commit
985685a
·
verified ·
1 Parent(s): e9b6e84

Upload folder using huggingface_hub

Browse files
Files changed (20) hide show
  1. artifacts/Hype1/results/acts/Qwen_Qwen2.5-7B-Instruct/layer10_n128_new256_maxlen512_states20000_seed42/aqua.npy +3 -0
  2. artifacts/Hype1/results/acts/Qwen_Qwen2.5-7B-Instruct/layer10_n128_new256_maxlen512_states20000_seed42/boolq.npy +3 -0
  3. artifacts/Hype1/results/acts/Qwen_Qwen2.5-7B-Instruct/layer10_n128_new256_maxlen512_states20000_seed42/collect.txt +30 -0
  4. artifacts/Hype1/results/acts/Qwen_Qwen2.5-7B-Instruct/layer10_n128_new256_maxlen512_states20000_seed42/commonsenseqa.npy +3 -0
  5. artifacts/Hype1/results/acts/Qwen_Qwen2.5-7B-Instruct/layer10_n128_new256_maxlen512_states20000_seed42/gsm8k.npy +3 -0
  6. artifacts/Hype1/results/acts/Qwen_Qwen2.5-7B-Instruct/layer10_n128_new256_maxlen512_states20000_seed42/meta.json +51 -0
  7. artifacts/Hype1/results/acts/Qwen_Qwen2.5-7B-Instruct/layer10_n128_new256_maxlen512_states20000_seed42/openbookqa.npy +3 -0
  8. artifacts/Hype1/results/acts/Qwen_Qwen2.5-7B-Instruct/layer10_n128_new256_maxlen512_states20000_seed42/piqa.npy +3 -0
  9. artifacts/Hype1/results/acts/Qwen_Qwen2.5-7B-Instruct/layer10_n128_new256_maxlen512_states20000_seed42/qasc.npy +3 -0
  10. artifacts/Hype1/results/acts/Qwen_Qwen2.5-7B-Instruct/layer10_n128_new256_maxlen512_states20000_seed42/strategyqa.npy +3 -0
  11. artifacts/Hype1/results/acts/meta-llama_Llama-2-7b-chat-hf/layer10_n128_new256_maxlen512_states20000_seed42/aqua.npy +3 -0
  12. artifacts/Hype1/results/acts/meta-llama_Llama-2-7b-chat-hf/layer10_n128_new256_maxlen512_states20000_seed42/boolq.npy +3 -0
  13. artifacts/Hype1/results/acts/meta-llama_Llama-2-7b-chat-hf/layer10_n128_new256_maxlen512_states20000_seed42/collect.txt +30 -0
  14. artifacts/Hype1/results/acts/meta-llama_Llama-2-7b-chat-hf/layer10_n128_new256_maxlen512_states20000_seed42/commonsenseqa.npy +3 -0
  15. artifacts/Hype1/results/acts/meta-llama_Llama-2-7b-chat-hf/layer10_n128_new256_maxlen512_states20000_seed42/gsm8k.npy +3 -0
  16. artifacts/Hype1/results/acts/meta-llama_Llama-2-7b-chat-hf/layer10_n128_new256_maxlen512_states20000_seed42/meta.json +51 -0
  17. artifacts/Hype1/results/acts/meta-llama_Llama-2-7b-chat-hf/layer10_n128_new256_maxlen512_states20000_seed42/openbookqa.npy +3 -0
  18. artifacts/Hype1/results/acts/meta-llama_Llama-2-7b-chat-hf/layer10_n128_new256_maxlen512_states20000_seed42/piqa.npy +3 -0
  19. artifacts/Hype1/results/acts/meta-llama_Llama-2-7b-chat-hf/layer10_n128_new256_maxlen512_states20000_seed42/qasc.npy +3 -0
  20. artifacts/Hype1/results/acts/meta-llama_Llama-2-7b-chat-hf/layer10_n128_new256_maxlen512_states20000_seed42/strategyqa.npy +3 -0
artifacts/Hype1/results/acts/Qwen_Qwen2.5-7B-Instruct/layer10_n128_new256_maxlen512_states20000_seed42/aqua.npy ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:c1c570097c0685a8fe3210e081d1e63c2dc621800427903ad9b42be043da76b6
3
+ size 286720128
artifacts/Hype1/results/acts/Qwen_Qwen2.5-7B-Instruct/layer10_n128_new256_maxlen512_states20000_seed42/boolq.npy ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:332b0c115c6a1997fa33dc126d367d2007585704eb94a25e0251d0180abf73f9
3
+ size 286720128
artifacts/Hype1/results/acts/Qwen_Qwen2.5-7B-Instruct/layer10_n128_new256_maxlen512_states20000_seed42/collect.txt ADDED
@@ -0,0 +1,30 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ [Cmd] collect_decode_acts.py --model Qwen/Qwen2.5-7B-Instruct --device cuda --model_dtype fp32 --layer 10 --n_prompts 128 --calib_max_new_tokens 256 --max_prompt_len 512 --per_task_max_states 20000 --batch_size 4 --seed 42 --out_dir results/acts/Qwen_Qwen2.5-7B-Instruct/layer10_n128_new256_maxlen512_states20000_seed42 --save_dtype fp32 --out_txt results/acts/Qwen_Qwen2.5-7B-Instruct/layer10_n128_new256_maxlen512_states20000_seed42/collect.txt --overwrite
2
+ [Data] tasks=['gsm8k', 'commonsenseqa', 'strategyqa', 'aqua', 'openbookqa', 'qasc', 'boolq', 'piqa']
3
+ [Data] task=gsm8k prompts=128
4
+ [Data] task=commonsenseqa prompts=128
5
+ [Data] task=strategyqa prompts=128
6
+ [Data] task=aqua prompts=128
7
+ [Data] task=openbookqa prompts=128
8
+ [Data] task=qasc prompts=128
9
+ [Data] task=boolq prompts=128
10
+ [Data] task=piqa prompts=128
11
+ [Collect] task=gsm8k
12
+ [Collect] task=commonsenseqa
13
+ [Collect] task=strategyqa
14
+ [Collect] task=aqua
15
+ [Collect] task=openbookqa
16
+ [Collect] task=qasc
17
+ [Collect] task=boolq
18
+ [Collect] task=piqa
19
+ [Collect] task=gsm8k raw_states=32768 x 3584
20
+ [Collect] task=commonsenseqa raw_states=32768 x 3584
21
+ [Collect] task=strategyqa raw_states=32768 x 3584
22
+ [Collect] task=aqua raw_states=32768 x 3584
23
+ [Collect] task=openbookqa raw_states=32768 x 3584
24
+ [Collect] task=qasc raw_states=32768 x 3584
25
+ [Collect] task=boolq raw_states=32768 x 3584
26
+ [Collect] task=piqa raw_states=32768 x 3584
27
+ [Fair] balanced states per task = 20000
28
+ [Done]
29
+ Saved meta: results/acts/Qwen_Qwen2.5-7B-Instruct/layer10_n128_new256_maxlen512_states20000_seed42/meta.json
30
+ Saved npy : results/acts/Qwen_Qwen2.5-7B-Instruct/layer10_n128_new256_maxlen512_states20000_seed42/*.npy
artifacts/Hype1/results/acts/Qwen_Qwen2.5-7B-Instruct/layer10_n128_new256_maxlen512_states20000_seed42/commonsenseqa.npy ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:6e19450db7ecaa63224cf42126e05d80e19a2d35a40d2c77a27cd015d539f716
3
+ size 286720128
artifacts/Hype1/results/acts/Qwen_Qwen2.5-7B-Instruct/layer10_n128_new256_maxlen512_states20000_seed42/gsm8k.npy ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:df7e4ccf8afafcf8efb989d498ca25abe0549ae917c6e61aa63b6cb593a5e231
3
+ size 286720128
artifacts/Hype1/results/acts/Qwen_Qwen2.5-7B-Instruct/layer10_n128_new256_maxlen512_states20000_seed42/meta.json ADDED
@@ -0,0 +1,51 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "config": {
3
+ "model": "Qwen/Qwen2.5-7B-Instruct",
4
+ "device": "cuda",
5
+ "model_dtype": "fp32",
6
+ "layer": 10,
7
+ "n_prompts": 128,
8
+ "max_prompt_len": 512,
9
+ "calib_max_new_tokens": 256,
10
+ "calib_decoding": "greedy",
11
+ "temperature": 0.7,
12
+ "top_p": 0.9,
13
+ "top_k": 0,
14
+ "batch_size": 4,
15
+ "per_task_max_states": 20000,
16
+ "balance_to": "min",
17
+ "seed": 42,
18
+ "save_dtype": "fp32"
19
+ },
20
+ "tasks": [
21
+ "gsm8k",
22
+ "commonsenseqa",
23
+ "strategyqa",
24
+ "aqua",
25
+ "openbookqa",
26
+ "qasc",
27
+ "boolq",
28
+ "piqa"
29
+ ],
30
+ "raw_counts": {
31
+ "gsm8k": 32768,
32
+ "commonsenseqa": 32768,
33
+ "strategyqa": 32768,
34
+ "aqua": 32768,
35
+ "openbookqa": 32768,
36
+ "qasc": 32768,
37
+ "boolq": 32768,
38
+ "piqa": 32768
39
+ },
40
+ "balanced_states_per_task": 20000,
41
+ "files": {
42
+ "gsm8k": "gsm8k.npy",
43
+ "commonsenseqa": "commonsenseqa.npy",
44
+ "strategyqa": "strategyqa.npy",
45
+ "aqua": "aqua.npy",
46
+ "openbookqa": "openbookqa.npy",
47
+ "qasc": "qasc.npy",
48
+ "boolq": "boolq.npy",
49
+ "piqa": "piqa.npy"
50
+ }
51
+ }
artifacts/Hype1/results/acts/Qwen_Qwen2.5-7B-Instruct/layer10_n128_new256_maxlen512_states20000_seed42/openbookqa.npy ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:c5586d1e23630b811935d72bb2e13540d5395292663aa12d2876d5e9471be179
3
+ size 286720128
artifacts/Hype1/results/acts/Qwen_Qwen2.5-7B-Instruct/layer10_n128_new256_maxlen512_states20000_seed42/piqa.npy ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:0ad5001af072f9ec2905e2b721423135bb2d11ba3ecbb45a0e0c50915bd47b8a
3
+ size 286720128
artifacts/Hype1/results/acts/Qwen_Qwen2.5-7B-Instruct/layer10_n128_new256_maxlen512_states20000_seed42/qasc.npy ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:4ba8669e27778d422a89839b2335090d54e6e9b8d551b78ee57dfa53b4782681
3
+ size 286720128
artifacts/Hype1/results/acts/Qwen_Qwen2.5-7B-Instruct/layer10_n128_new256_maxlen512_states20000_seed42/strategyqa.npy ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:ad7406cf57b0409cc764e69dac4a59958496fba917fff3314fe3a3a3d8af944f
3
+ size 286720128
artifacts/Hype1/results/acts/meta-llama_Llama-2-7b-chat-hf/layer10_n128_new256_maxlen512_states20000_seed42/aqua.npy ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:f298d287b81536c351a9ed4a89dc6dd7b291514ae9e2ff4171993b55eca01467
3
+ size 304087168
artifacts/Hype1/results/acts/meta-llama_Llama-2-7b-chat-hf/layer10_n128_new256_maxlen512_states20000_seed42/boolq.npy ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:65a25e641a96f1e1b1e5a0d4727c1996b0c3f7770d1fc13f79f5f8c0022d0da0
3
+ size 304087168
artifacts/Hype1/results/acts/meta-llama_Llama-2-7b-chat-hf/layer10_n128_new256_maxlen512_states20000_seed42/collect.txt ADDED
@@ -0,0 +1,30 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ [Cmd] collect_decode_acts.py --model meta-llama/Llama-2-7b-chat-hf --device cuda --model_dtype fp32 --layer 10 --n_prompts 128 --calib_max_new_tokens 256 --max_prompt_len 512 --per_task_max_states 20000 --batch_size 4 --seed 42 --out_dir results/acts/meta-llama_Llama-2-7b-chat-hf/layer10_n128_new256_maxlen512_states20000_seed42 --save_dtype fp32 --out_txt results/acts/meta-llama_Llama-2-7b-chat-hf/layer10_n128_new256_maxlen512_states20000_seed42/collect.txt --overwrite
2
+ [Data] tasks=['gsm8k', 'commonsenseqa', 'strategyqa', 'aqua', 'openbookqa', 'qasc', 'boolq', 'piqa']
3
+ [Data] task=gsm8k prompts=128
4
+ [Data] task=commonsenseqa prompts=128
5
+ [Data] task=strategyqa prompts=128
6
+ [Data] task=aqua prompts=128
7
+ [Data] task=openbookqa prompts=128
8
+ [Data] task=qasc prompts=128
9
+ [Data] task=boolq prompts=128
10
+ [Data] task=piqa prompts=128
11
+ [Collect] task=gsm8k
12
+ [Collect] task=commonsenseqa
13
+ [Collect] task=strategyqa
14
+ [Collect] task=aqua
15
+ [Collect] task=openbookqa
16
+ [Collect] task=qasc
17
+ [Collect] task=boolq
18
+ [Collect] task=piqa
19
+ [Collect] task=gsm8k raw_states=18560 x 4096
20
+ [Collect] task=commonsenseqa raw_states=21001 x 4096
21
+ [Collect] task=strategyqa raw_states=23509 x 4096
22
+ [Collect] task=aqua raw_states=24667 x 4096
23
+ [Collect] task=openbookqa raw_states=20494 x 4096
24
+ [Collect] task=qasc raw_states=25433 x 4096
25
+ [Collect] task=boolq raw_states=22885 x 4096
26
+ [Collect] task=piqa raw_states=21000 x 4096
27
+ [Fair] balanced states per task = 18560
28
+ [Done]
29
+ Saved meta: results/acts/meta-llama_Llama-2-7b-chat-hf/layer10_n128_new256_maxlen512_states20000_seed42/meta.json
30
+ Saved npy : results/acts/meta-llama_Llama-2-7b-chat-hf/layer10_n128_new256_maxlen512_states20000_seed42/*.npy
artifacts/Hype1/results/acts/meta-llama_Llama-2-7b-chat-hf/layer10_n128_new256_maxlen512_states20000_seed42/commonsenseqa.npy ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:604bf6b8d97ad18b806f92a90e3961f176ae11339b3e8586414bac04f0d1f7f9
3
+ size 304087168
artifacts/Hype1/results/acts/meta-llama_Llama-2-7b-chat-hf/layer10_n128_new256_maxlen512_states20000_seed42/gsm8k.npy ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:289c13700e324287c1d6a7dfbb0d2a1fa04a552a142bd6e84fa21708043e1925
3
+ size 304087168
artifacts/Hype1/results/acts/meta-llama_Llama-2-7b-chat-hf/layer10_n128_new256_maxlen512_states20000_seed42/meta.json ADDED
@@ -0,0 +1,51 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "config": {
3
+ "model": "meta-llama/Llama-2-7b-chat-hf",
4
+ "device": "cuda",
5
+ "model_dtype": "fp32",
6
+ "layer": 10,
7
+ "n_prompts": 128,
8
+ "max_prompt_len": 512,
9
+ "calib_max_new_tokens": 256,
10
+ "calib_decoding": "greedy",
11
+ "temperature": 0.7,
12
+ "top_p": 0.9,
13
+ "top_k": 0,
14
+ "batch_size": 4,
15
+ "per_task_max_states": 20000,
16
+ "balance_to": "min",
17
+ "seed": 42,
18
+ "save_dtype": "fp32"
19
+ },
20
+ "tasks": [
21
+ "gsm8k",
22
+ "commonsenseqa",
23
+ "strategyqa",
24
+ "aqua",
25
+ "openbookqa",
26
+ "qasc",
27
+ "boolq",
28
+ "piqa"
29
+ ],
30
+ "raw_counts": {
31
+ "gsm8k": 18560,
32
+ "commonsenseqa": 21001,
33
+ "strategyqa": 23509,
34
+ "aqua": 24667,
35
+ "openbookqa": 20494,
36
+ "qasc": 25433,
37
+ "boolq": 22885,
38
+ "piqa": 21000
39
+ },
40
+ "balanced_states_per_task": 18560,
41
+ "files": {
42
+ "gsm8k": "gsm8k.npy",
43
+ "commonsenseqa": "commonsenseqa.npy",
44
+ "strategyqa": "strategyqa.npy",
45
+ "aqua": "aqua.npy",
46
+ "openbookqa": "openbookqa.npy",
47
+ "qasc": "qasc.npy",
48
+ "boolq": "boolq.npy",
49
+ "piqa": "piqa.npy"
50
+ }
51
+ }
artifacts/Hype1/results/acts/meta-llama_Llama-2-7b-chat-hf/layer10_n128_new256_maxlen512_states20000_seed42/openbookqa.npy ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:8a5b094d1efc9c7929d4b218ea6bfbe4df3b072d26ecefb1a1225e97681c74ef
3
+ size 304087168
artifacts/Hype1/results/acts/meta-llama_Llama-2-7b-chat-hf/layer10_n128_new256_maxlen512_states20000_seed42/piqa.npy ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:bb16b852f8e33147c56c183f6e3a2e23ce9aac8403fc6e8d30fbe0bb3988341e
3
+ size 304087168
artifacts/Hype1/results/acts/meta-llama_Llama-2-7b-chat-hf/layer10_n128_new256_maxlen512_states20000_seed42/qasc.npy ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:248557735b4175fa06f50ce2370884f2cb8785a6c4c0e977db619ff29735b24e
3
+ size 304087168
artifacts/Hype1/results/acts/meta-llama_Llama-2-7b-chat-hf/layer10_n128_new256_maxlen512_states20000_seed42/strategyqa.npy ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:8b5635ce8b379d9d9b88d17b10979dd9d54e25ec288d74082db1c54c0a4b05d7
3
+ size 304087168