forked from thudzj/ScaledRoPE
-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathlog.txt
52 lines (44 loc) · 3.16 KB
/
log.txt
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
pip install fastchat
cd LongChat
pip install -e .
cd longeval
# baseline vicuna without extended context
python3 eval.py --model-name-or-path lmsys/vicuna-7b-v1.3 --task lines --eval_shortest_only
accuracy: 0.0
# baseline vicuna wit linearly-extended context
python3 eval.py --model-name-or-path lmsys/vicuna-7b-v1.3 --task lines --eval_shortest_only --longchat_ratio 4 --interpolation_type linear
accuracy: 0.0
# baseline vicuna wit ntk-extended context
python3 eval.py --model-name-or-path lmsys/vicuna-7b-v1.3 --task lines --eval_shortest_only --longchat_ratio 4 --interpolation_type ntk
accuracy: 0.38
tensor([1.0000e+00, 8.4712e-01, 7.1761e-01, 6.0790e-01, 5.1496e-01, 4.3623e-01,
3.6954e-01, 3.1304e-01, 2.6518e-01, 2.2464e-01, 1.9030e-01, 1.6120e-01,
1.3656e-01, 1.1568e-01, 9.7996e-02, 8.3014e-02, 7.0323e-02, 5.9572e-02,
5.0464e-02, 4.2749e-02, 3.6213e-02, 3.0677e-02, 2.5987e-02, 2.2014e-02,
1.8648e-02, 1.5797e-02, 1.3382e-02, 1.1336e-02, 9.6032e-03, 8.1351e-03,
6.8914e-03, 5.8378e-03, 4.9453e-03, 4.1892e-03, 3.5488e-03, 3.0062e-03,
2.5466e-03, 2.1573e-03, 1.8275e-03, 1.5481e-03, 1.3114e-03, 1.1109e-03,
9.4108e-04, 7.9721e-04, 6.7533e-04, 5.7208e-04, 4.8462e-04, 4.1053e-04,
3.4777e-04, 2.9460e-04, 2.4956e-04, 2.1141e-04, 1.7909e-04, 1.5171e-04,
1.2851e-04, 1.0887e-04, 9.2222e-05, 7.8123e-05, 6.6179e-05, 5.6062e-05,
4.7491e-05, 4.0230e-05, 3.4080e-05, 2.8870e-05], dtype=torch.float32)
# our
python3 eval.py --model-name-or-path lmsys/vicuna-7b-v1.3 --task lines --eval_shortest_only --longchat_ratio 4 --interpolation_type our
accuracy: 0.38
tensor([1.0000e+00, 8.4712e-01, 7.1761e-01, 6.0790e-01, 5.1496e-01, 4.3623e-01,
3.6954e-01, 3.1304e-01, 2.6518e-01, 2.2464e-01, 1.9030e-01, 1.6120e-01,
1.3656e-01, 1.1568e-01, 9.7996e-02, 8.3014e-02, 7.0323e-02, 5.9572e-02,
5.0464e-02, 4.2749e-02, 3.6213e-02, 3.0677e-02, 2.5987e-02, 2.2014e-02,
1.8648e-02, 1.5797e-02, 1.3382e-02, 1.1336e-02, 9.6032e-03, 8.1351e-03,
6.8914e-03, 5.8378e-03, 4.9453e-03, 4.1892e-03, 3.5488e-03, 3.0062e-03,
2.5466e-03, 2.1573e-03, 1.8275e-03, 1.5481e-03, 1.3114e-03, 1.1109e-03,
9.4108e-04, 7.9721e-04, 6.7533e-04, 5.7208e-04, 4.8462e-04, 4.1053e-04,
3.4777e-04, 2.9460e-04, 2.4956e-04, 2.1141e-04, 1.7909e-04, 1.5171e-04,
1.2851e-04, 1.0887e-04, 9.2222e-05, 7.8123e-05, 6.6179e-05, 5.6062e-05,
4.7491e-05, 4.0230e-05, 3.4080e-05, 2.8870e-05], dtype=torch.float32)
------------------ensemble------------------
python3 eval.py --model-name-or-path lmsys/vicuna-7b-v1.3 --task lines --eval_shortest_only --longchat_ratio 4 --interpolation_type ntk --longchat_flash_attn
CUDA_VISIBLE_DEVICES=4 python3 eval.py --model-name-or-path daryl149/Llama-2-7b-chat-hf --task lines --eval_shortest_only --longchat_ratio 4 --interpolation_type ntk --longchat_flash_attn --random_ratios 1.5 2
-----------mmlu----------
CUDA_VISIBLE_DEVICES=4 python3 eval.py --model-name-or-path daryl149/Llama-2-7b-chat-hf --task lines --eval_shortest_only --longchat_ratio 4 --interpolation_type ntk --longchat_flash_attn --random_ratios 1
Average accuracy: 0.458