Upload folder using huggingface_hub
Browse filesThis view is limited to 50 files because it contains too many changes.
See raw diff
- .gitattributes +349 -0
- README.md +16 -0
- activation.py +101 -0
- activation_all.log +0 -0
- activation_all.py +192 -0
- activation_all.sh +16 -0
- activation_llama.py +101 -0
- activation_llama_all.py +146 -0
- activation_qwen.py +109 -0
- activation_qwen_all.py +126 -0
- activation_single.py +76 -0
- activation_single.sh +55 -0
- activations-old/README.md +1 -0
- activations-old/qwen2.5-0.5b_english_wiki_1.5B_basque_corpus/activation.en.train.qwen-checkpoint-1200 +3 -0
- activations-old/qwen2.5-0.5b_english_wiki_1.5B_basque_corpus/activation.en.train.qwen-checkpoint-1500 +3 -0
- activations-old/qwen2.5-0.5b_english_wiki_1.5B_basque_corpus/activation.en.train.qwen-checkpoint-1800 +3 -0
- activations-old/qwen2.5-0.5b_english_wiki_1.5B_basque_corpus/activation.en.train.qwen-checkpoint-2100 +3 -0
- activations-old/qwen2.5-0.5b_english_wiki_1.5B_basque_corpus/activation.en.train.qwen-checkpoint-2400 +3 -0
- activations-old/qwen2.5-0.5b_english_wiki_1.5B_basque_corpus/activation.en.train.qwen-checkpoint-2700 +3 -0
- activations-old/qwen2.5-0.5b_english_wiki_1.5B_basque_corpus/activation.en.train.qwen-checkpoint-300 +3 -0
- activations-old/qwen2.5-0.5b_english_wiki_1.5B_basque_corpus/activation.en.train.qwen-checkpoint-3000 +3 -0
- activations-old/qwen2.5-0.5b_english_wiki_1.5B_basque_corpus/activation.en.train.qwen-checkpoint-3231 +3 -0
- activations-old/qwen2.5-0.5b_english_wiki_1.5B_basque_corpus/activation.en.train.qwen-checkpoint-600 +3 -0
- activations-old/qwen2.5-0.5b_english_wiki_1.5B_basque_corpus/activation.en.train.qwen-checkpoint-900 +3 -0
- activations-old/qwen2.5-0.5b_english_wiki_1.5B_basque_corpus/activation.eu.train.qwen-checkpoint-1200 +3 -0
- activations-old/qwen2.5-0.5b_english_wiki_1.5B_basque_corpus/activation.eu.train.qwen-checkpoint-1500 +3 -0
- activations-old/qwen2.5-0.5b_english_wiki_1.5B_basque_corpus/activation.eu.train.qwen-checkpoint-1800 +3 -0
- activations-old/qwen2.5-0.5b_english_wiki_1.5B_basque_corpus/activation.eu.train.qwen-checkpoint-2100 +3 -0
- activations-old/qwen2.5-0.5b_english_wiki_1.5B_basque_corpus/activation.eu.train.qwen-checkpoint-2400 +3 -0
- activations-old/qwen2.5-0.5b_english_wiki_1.5B_basque_corpus/activation.eu.train.qwen-checkpoint-2700 +3 -0
- activations-old/qwen2.5-0.5b_english_wiki_1.5B_basque_corpus/activation.eu.train.qwen-checkpoint-300 +3 -0
- activations-old/qwen2.5-0.5b_english_wiki_1.5B_basque_corpus/activation.eu.train.qwen-checkpoint-3000 +3 -0
- activations-old/qwen2.5-0.5b_english_wiki_1.5B_basque_corpus/activation.eu.train.qwen-checkpoint-3231 +3 -0
- activations-old/qwen2.5-0.5b_english_wiki_1.5B_basque_corpus/activation.eu.train.qwen-checkpoint-600 +3 -0
- activations-old/qwen2.5-0.5b_english_wiki_1.5B_basque_corpus/activation.eu.train.qwen-checkpoint-900 +3 -0
- activations-old/qwen2.5-0.5b_english_wiki_300M_en_1.5Bbasque_corpus/activation.en.train.qwen-checkpoint-1200 +3 -0
- activations-old/qwen2.5-0.5b_english_wiki_300M_en_1.5Bbasque_corpus/activation.en.train.qwen-checkpoint-1500 +3 -0
- activations-old/qwen2.5-0.5b_english_wiki_300M_en_1.5Bbasque_corpus/activation.en.train.qwen-checkpoint-1800 +3 -0
- activations-old/qwen2.5-0.5b_english_wiki_300M_en_1.5Bbasque_corpus/activation.en.train.qwen-checkpoint-2100 +3 -0
- activations-old/qwen2.5-0.5b_english_wiki_300M_en_1.5Bbasque_corpus/activation.en.train.qwen-checkpoint-2400 +3 -0
- activations-old/qwen2.5-0.5b_english_wiki_300M_en_1.5Bbasque_corpus/activation.en.train.qwen-checkpoint-2700 +3 -0
- activations-old/qwen2.5-0.5b_english_wiki_300M_en_1.5Bbasque_corpus/activation.en.train.qwen-checkpoint-2972 +3 -0
- activations-old/qwen2.5-0.5b_english_wiki_300M_en_1.5Bbasque_corpus/activation.en.train.qwen-checkpoint-300 +3 -0
- activations-old/qwen2.5-0.5b_english_wiki_300M_en_1.5Bbasque_corpus/activation.en.train.qwen-checkpoint-600 +3 -0
- activations-old/qwen2.5-0.5b_english_wiki_300M_en_1.5Bbasque_corpus/activation.en.train.qwen-checkpoint-900 +3 -0
- activations-old/qwen2.5-0.5b_english_wiki_300M_en_1.5Bbasque_corpus/activation.eu.train.qwen-checkpoint-1200 +3 -0
- activations-old/qwen2.5-0.5b_english_wiki_300M_en_1.5Bbasque_corpus/activation.eu.train.qwen-checkpoint-1500 +3 -0
- activations-old/qwen2.5-0.5b_english_wiki_300M_en_1.5Bbasque_corpus/activation.eu.train.qwen-checkpoint-1800 +3 -0
- activations-old/qwen2.5-0.5b_english_wiki_300M_en_1.5Bbasque_corpus/activation.eu.train.qwen-checkpoint-2100 +3 -0
- activations-old/qwen2.5-0.5b_english_wiki_300M_en_1.5Bbasque_corpus/activation.eu.train.qwen-checkpoint-2400 +3 -0
.gitattributes
CHANGED
|
@@ -33,3 +33,352 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
|
|
| 33 |
*.zip filter=lfs diff=lfs merge=lfs -text
|
| 34 |
*.zst filter=lfs diff=lfs merge=lfs -text
|
| 35 |
*tfevents* filter=lfs diff=lfs merge=lfs -text
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 33 |
*.zip filter=lfs diff=lfs merge=lfs -text
|
| 34 |
*.zst filter=lfs diff=lfs merge=lfs -text
|
| 35 |
*tfevents* filter=lfs diff=lfs merge=lfs -text
|
| 36 |
+
activations/qwen2.5-0.5b_english_wiki_1.5B_basque_corpus/activation.en.train.qwen-checkpoint-1200 filter=lfs diff=lfs merge=lfs -text
|
| 37 |
+
activations/qwen2.5-0.5b_english_wiki_1.5B_basque_corpus/activation.en.train.qwen-checkpoint-1500 filter=lfs diff=lfs merge=lfs -text
|
| 38 |
+
activations/qwen2.5-0.5b_english_wiki_1.5B_basque_corpus/activation.en.train.qwen-checkpoint-1800 filter=lfs diff=lfs merge=lfs -text
|
| 39 |
+
activations/qwen2.5-0.5b_english_wiki_1.5B_basque_corpus/activation.en.train.qwen-checkpoint-2100 filter=lfs diff=lfs merge=lfs -text
|
| 40 |
+
activations/qwen2.5-0.5b_english_wiki_1.5B_basque_corpus/activation.en.train.qwen-checkpoint-2400 filter=lfs diff=lfs merge=lfs -text
|
| 41 |
+
activations/qwen2.5-0.5b_english_wiki_1.5B_basque_corpus/activation.en.train.qwen-checkpoint-2700 filter=lfs diff=lfs merge=lfs -text
|
| 42 |
+
activations/qwen2.5-0.5b_english_wiki_1.5B_basque_corpus/activation.en.train.qwen-checkpoint-300 filter=lfs diff=lfs merge=lfs -text
|
| 43 |
+
activations/qwen2.5-0.5b_english_wiki_1.5B_basque_corpus/activation.en.train.qwen-checkpoint-3000 filter=lfs diff=lfs merge=lfs -text
|
| 44 |
+
activations/qwen2.5-0.5b_english_wiki_1.5B_basque_corpus/activation.en.train.qwen-checkpoint-3231 filter=lfs diff=lfs merge=lfs -text
|
| 45 |
+
activations/qwen2.5-0.5b_english_wiki_1.5B_basque_corpus/activation.en.train.qwen-checkpoint-600 filter=lfs diff=lfs merge=lfs -text
|
| 46 |
+
activations/qwen2.5-0.5b_english_wiki_1.5B_basque_corpus/activation.en.train.qwen-checkpoint-900 filter=lfs diff=lfs merge=lfs -text
|
| 47 |
+
activations/qwen2.5-0.5b_english_wiki_1.5B_basque_corpus/activation.eu.train.qwen-checkpoint-1200 filter=lfs diff=lfs merge=lfs -text
|
| 48 |
+
activations/qwen2.5-0.5b_english_wiki_1.5B_basque_corpus/activation.eu.train.qwen-checkpoint-1500 filter=lfs diff=lfs merge=lfs -text
|
| 49 |
+
activations/qwen2.5-0.5b_english_wiki_1.5B_basque_corpus/activation.eu.train.qwen-checkpoint-1800 filter=lfs diff=lfs merge=lfs -text
|
| 50 |
+
activations/qwen2.5-0.5b_english_wiki_1.5B_basque_corpus/activation.eu.train.qwen-checkpoint-2100 filter=lfs diff=lfs merge=lfs -text
|
| 51 |
+
activations/qwen2.5-0.5b_english_wiki_1.5B_basque_corpus/activation.eu.train.qwen-checkpoint-2400 filter=lfs diff=lfs merge=lfs -text
|
| 52 |
+
activations/qwen2.5-0.5b_english_wiki_1.5B_basque_corpus/activation.eu.train.qwen-checkpoint-2700 filter=lfs diff=lfs merge=lfs -text
|
| 53 |
+
activations/qwen2.5-0.5b_english_wiki_1.5B_basque_corpus/activation.eu.train.qwen-checkpoint-300 filter=lfs diff=lfs merge=lfs -text
|
| 54 |
+
activations/qwen2.5-0.5b_english_wiki_1.5B_basque_corpus/activation.eu.train.qwen-checkpoint-3000 filter=lfs diff=lfs merge=lfs -text
|
| 55 |
+
activations/qwen2.5-0.5b_english_wiki_1.5B_basque_corpus/activation.eu.train.qwen-checkpoint-3231 filter=lfs diff=lfs merge=lfs -text
|
| 56 |
+
activations/qwen2.5-0.5b_english_wiki_1.5B_basque_corpus/activation.eu.train.qwen-checkpoint-600 filter=lfs diff=lfs merge=lfs -text
|
| 57 |
+
activations/qwen2.5-0.5b_english_wiki_1.5B_basque_corpus/activation.eu.train.qwen-checkpoint-900 filter=lfs diff=lfs merge=lfs -text
|
| 58 |
+
activations/qwen2.5-0.5b_english_wiki_300M_en_1.5Bbasque_corpus/activation.en.train.qwen-checkpoint-1200 filter=lfs diff=lfs merge=lfs -text
|
| 59 |
+
activations/qwen2.5-0.5b_english_wiki_300M_en_1.5Bbasque_corpus/activation.en.train.qwen-checkpoint-1500 filter=lfs diff=lfs merge=lfs -text
|
| 60 |
+
activations/qwen2.5-0.5b_english_wiki_300M_en_1.5Bbasque_corpus/activation.en.train.qwen-checkpoint-1800 filter=lfs diff=lfs merge=lfs -text
|
| 61 |
+
activations/qwen2.5-0.5b_english_wiki_300M_en_1.5Bbasque_corpus/activation.en.train.qwen-checkpoint-2100 filter=lfs diff=lfs merge=lfs -text
|
| 62 |
+
activations/qwen2.5-0.5b_english_wiki_300M_en_1.5Bbasque_corpus/activation.en.train.qwen-checkpoint-2400 filter=lfs diff=lfs merge=lfs -text
|
| 63 |
+
activations/qwen2.5-0.5b_english_wiki_300M_en_1.5Bbasque_corpus/activation.en.train.qwen-checkpoint-2700 filter=lfs diff=lfs merge=lfs -text
|
| 64 |
+
activations/qwen2.5-0.5b_english_wiki_300M_en_1.5Bbasque_corpus/activation.en.train.qwen-checkpoint-2972 filter=lfs diff=lfs merge=lfs -text
|
| 65 |
+
activations/qwen2.5-0.5b_english_wiki_300M_en_1.5Bbasque_corpus/activation.en.train.qwen-checkpoint-300 filter=lfs diff=lfs merge=lfs -text
|
| 66 |
+
activations/qwen2.5-0.5b_english_wiki_300M_en_1.5Bbasque_corpus/activation.en.train.qwen-checkpoint-600 filter=lfs diff=lfs merge=lfs -text
|
| 67 |
+
activations/qwen2.5-0.5b_english_wiki_300M_en_1.5Bbasque_corpus/activation.en.train.qwen-checkpoint-900 filter=lfs diff=lfs merge=lfs -text
|
| 68 |
+
activations/qwen2.5-0.5b_english_wiki_300M_en_1.5Bbasque_corpus/activation.eu.train.qwen-checkpoint-1200 filter=lfs diff=lfs merge=lfs -text
|
| 69 |
+
activations/qwen2.5-0.5b_english_wiki_300M_en_1.5Bbasque_corpus/activation.eu.train.qwen-checkpoint-1500 filter=lfs diff=lfs merge=lfs -text
|
| 70 |
+
activations/qwen2.5-0.5b_english_wiki_300M_en_1.5Bbasque_corpus/activation.eu.train.qwen-checkpoint-1800 filter=lfs diff=lfs merge=lfs -text
|
| 71 |
+
activations/qwen2.5-0.5b_english_wiki_300M_en_1.5Bbasque_corpus/activation.eu.train.qwen-checkpoint-2100 filter=lfs diff=lfs merge=lfs -text
|
| 72 |
+
activations/qwen2.5-0.5b_english_wiki_300M_en_1.5Bbasque_corpus/activation.eu.train.qwen-checkpoint-2400 filter=lfs diff=lfs merge=lfs -text
|
| 73 |
+
activations/qwen2.5-0.5b_english_wiki_300M_en_1.5Bbasque_corpus/activation.eu.train.qwen-checkpoint-2700 filter=lfs diff=lfs merge=lfs -text
|
| 74 |
+
activations/qwen2.5-0.5b_english_wiki_300M_en_1.5Bbasque_corpus/activation.eu.train.qwen-checkpoint-2972 filter=lfs diff=lfs merge=lfs -text
|
| 75 |
+
activations/qwen2.5-0.5b_english_wiki_300M_en_1.5Bbasque_corpus/activation.eu.train.qwen-checkpoint-300 filter=lfs diff=lfs merge=lfs -text
|
| 76 |
+
activations/qwen2.5-0.5b_english_wiki_300M_en_1.5Bbasque_corpus/activation.eu.train.qwen-checkpoint-600 filter=lfs diff=lfs merge=lfs -text
|
| 77 |
+
activations/qwen2.5-0.5b_english_wiki_300M_en_1.5Bbasque_corpus/activation.eu.train.qwen-checkpoint-900 filter=lfs diff=lfs merge=lfs -text
|
| 78 |
+
activations/qwen2.5-0.5b_english_wiki_300M_en_1.5Birish_corpus/activation.en.train.qwen-checkpoint-1200 filter=lfs diff=lfs merge=lfs -text
|
| 79 |
+
activations/qwen2.5-0.5b_english_wiki_300M_en_1.5Birish_corpus/activation.en.train.qwen-checkpoint-1500 filter=lfs diff=lfs merge=lfs -text
|
| 80 |
+
activations/qwen2.5-0.5b_english_wiki_300M_en_1.5Birish_corpus/activation.en.train.qwen-checkpoint-1800 filter=lfs diff=lfs merge=lfs -text
|
| 81 |
+
activations/qwen2.5-0.5b_english_wiki_300M_en_1.5Birish_corpus/activation.en.train.qwen-checkpoint-2100 filter=lfs diff=lfs merge=lfs -text
|
| 82 |
+
activations/qwen2.5-0.5b_english_wiki_300M_en_1.5Birish_corpus/activation.en.train.qwen-checkpoint-2400 filter=lfs diff=lfs merge=lfs -text
|
| 83 |
+
activations/qwen2.5-0.5b_english_wiki_300M_en_1.5Birish_corpus/activation.en.train.qwen-checkpoint-2700 filter=lfs diff=lfs merge=lfs -text
|
| 84 |
+
activations/qwen2.5-0.5b_english_wiki_300M_en_1.5Birish_corpus/activation.en.train.qwen-checkpoint-300 filter=lfs diff=lfs merge=lfs -text
|
| 85 |
+
activations/qwen2.5-0.5b_english_wiki_300M_en_1.5Birish_corpus/activation.en.train.qwen-checkpoint-3000 filter=lfs diff=lfs merge=lfs -text
|
| 86 |
+
activations/qwen2.5-0.5b_english_wiki_300M_en_1.5Birish_corpus/activation.en.train.qwen-checkpoint-3300 filter=lfs diff=lfs merge=lfs -text
|
| 87 |
+
activations/qwen2.5-0.5b_english_wiki_300M_en_1.5Birish_corpus/activation.en.train.qwen-checkpoint-3600 filter=lfs diff=lfs merge=lfs -text
|
| 88 |
+
activations/qwen2.5-0.5b_english_wiki_300M_en_1.5Birish_corpus/activation.en.train.qwen-checkpoint-3900 filter=lfs diff=lfs merge=lfs -text
|
| 89 |
+
activations/qwen2.5-0.5b_english_wiki_300M_en_1.5Birish_corpus/activation.en.train.qwen-checkpoint-4200 filter=lfs diff=lfs merge=lfs -text
|
| 90 |
+
activations/qwen2.5-0.5b_english_wiki_300M_en_1.5Birish_corpus/activation.en.train.qwen-checkpoint-4500 filter=lfs diff=lfs merge=lfs -text
|
| 91 |
+
activations/qwen2.5-0.5b_english_wiki_300M_en_1.5Birish_corpus/activation.en.train.qwen-checkpoint-4800 filter=lfs diff=lfs merge=lfs -text
|
| 92 |
+
activations/qwen2.5-0.5b_english_wiki_300M_en_1.5Birish_corpus/activation.en.train.qwen-checkpoint-4888 filter=lfs diff=lfs merge=lfs -text
|
| 93 |
+
activations/qwen2.5-0.5b_english_wiki_300M_en_1.5Birish_corpus/activation.en.train.qwen-checkpoint-600 filter=lfs diff=lfs merge=lfs -text
|
| 94 |
+
activations/qwen2.5-0.5b_english_wiki_300M_en_1.5Birish_corpus/activation.en.train.qwen-checkpoint-900 filter=lfs diff=lfs merge=lfs -text
|
| 95 |
+
activations/qwen2.5-0.5b_english_wiki_300M_en_1.5Birish_corpus/activation.ga.train.qwen-checkpoint-1200 filter=lfs diff=lfs merge=lfs -text
|
| 96 |
+
activations/qwen2.5-0.5b_english_wiki_300M_en_1.5Birish_corpus/activation.ga.train.qwen-checkpoint-1500 filter=lfs diff=lfs merge=lfs -text
|
| 97 |
+
activations/qwen2.5-0.5b_english_wiki_300M_en_1.5Birish_corpus/activation.ga.train.qwen-checkpoint-1800 filter=lfs diff=lfs merge=lfs -text
|
| 98 |
+
activations/qwen2.5-0.5b_english_wiki_300M_en_1.5Birish_corpus/activation.ga.train.qwen-checkpoint-2100 filter=lfs diff=lfs merge=lfs -text
|
| 99 |
+
activations/qwen2.5-0.5b_english_wiki_300M_en_1.5Birish_corpus/activation.ga.train.qwen-checkpoint-2400 filter=lfs diff=lfs merge=lfs -text
|
| 100 |
+
activations/qwen2.5-0.5b_english_wiki_300M_en_1.5Birish_corpus/activation.ga.train.qwen-checkpoint-2700 filter=lfs diff=lfs merge=lfs -text
|
| 101 |
+
activations/qwen2.5-0.5b_english_wiki_300M_en_1.5Birish_corpus/activation.ga.train.qwen-checkpoint-300 filter=lfs diff=lfs merge=lfs -text
|
| 102 |
+
activations/qwen2.5-0.5b_english_wiki_300M_en_1.5Birish_corpus/activation.ga.train.qwen-checkpoint-3000 filter=lfs diff=lfs merge=lfs -text
|
| 103 |
+
activations/qwen2.5-0.5b_english_wiki_300M_en_1.5Birish_corpus/activation.ga.train.qwen-checkpoint-3300 filter=lfs diff=lfs merge=lfs -text
|
| 104 |
+
activations/qwen2.5-0.5b_english_wiki_300M_en_1.5Birish_corpus/activation.ga.train.qwen-checkpoint-3600 filter=lfs diff=lfs merge=lfs -text
|
| 105 |
+
activations/qwen2.5-0.5b_english_wiki_300M_en_1.5Birish_corpus/activation.ga.train.qwen-checkpoint-3900 filter=lfs diff=lfs merge=lfs -text
|
| 106 |
+
activations/qwen2.5-0.5b_english_wiki_300M_en_1.5Birish_corpus/activation.ga.train.qwen-checkpoint-4200 filter=lfs diff=lfs merge=lfs -text
|
| 107 |
+
activations/qwen2.5-0.5b_english_wiki_300M_en_1.5Birish_corpus/activation.ga.train.qwen-checkpoint-4500 filter=lfs diff=lfs merge=lfs -text
|
| 108 |
+
activations/qwen2.5-0.5b_english_wiki_300M_en_1.5Birish_corpus/activation.ga.train.qwen-checkpoint-4800 filter=lfs diff=lfs merge=lfs -text
|
| 109 |
+
activations/qwen2.5-0.5b_english_wiki_300M_en_1.5Birish_corpus/activation.ga.train.qwen-checkpoint-4888 filter=lfs diff=lfs merge=lfs -text
|
| 110 |
+
activations/qwen2.5-0.5b_english_wiki_300M_en_1.5Birish_corpus/activation.ga.train.qwen-checkpoint-600 filter=lfs diff=lfs merge=lfs -text
|
| 111 |
+
activations/qwen2.5-0.5b_english_wiki_300M_en_1.5Birish_corpus/activation.ga.train.qwen-checkpoint-900 filter=lfs diff=lfs merge=lfs -text
|
| 112 |
+
activations/qwen2.5-0.5b_english_wiki_750M_chinese_wikipedia_corpus/activation.en.train.qwen-checkpoint-1200 filter=lfs diff=lfs merge=lfs -text
|
| 113 |
+
activations/qwen2.5-0.5b_english_wiki_750M_chinese_wikipedia_corpus/activation.en.train.qwen-checkpoint-1500 filter=lfs diff=lfs merge=lfs -text
|
| 114 |
+
activations/qwen2.5-0.5b_english_wiki_750M_chinese_wikipedia_corpus/activation.en.train.qwen-checkpoint-1800 filter=lfs diff=lfs merge=lfs -text
|
| 115 |
+
activations/qwen2.5-0.5b_english_wiki_750M_chinese_wikipedia_corpus/activation.en.train.qwen-checkpoint-1962 filter=lfs diff=lfs merge=lfs -text
|
| 116 |
+
activations/qwen2.5-0.5b_english_wiki_750M_chinese_wikipedia_corpus/activation.en.train.qwen-checkpoint-300 filter=lfs diff=lfs merge=lfs -text
|
| 117 |
+
activations/qwen2.5-0.5b_english_wiki_750M_chinese_wikipedia_corpus/activation.en.train.qwen-checkpoint-600 filter=lfs diff=lfs merge=lfs -text
|
| 118 |
+
activations/qwen2.5-0.5b_english_wiki_750M_chinese_wikipedia_corpus/activation.en.train.qwen-checkpoint-900 filter=lfs diff=lfs merge=lfs -text
|
| 119 |
+
activations/qwen2.5-0.5b_english_wiki_750M_chinese_wikipedia_corpus/activation.zh.train.qwen-checkpoint-1200 filter=lfs diff=lfs merge=lfs -text
|
| 120 |
+
activations/qwen2.5-0.5b_english_wiki_750M_chinese_wikipedia_corpus/activation.zh.train.qwen-checkpoint-1500 filter=lfs diff=lfs merge=lfs -text
|
| 121 |
+
activations/qwen2.5-0.5b_english_wiki_750M_chinese_wikipedia_corpus/activation.zh.train.qwen-checkpoint-1800 filter=lfs diff=lfs merge=lfs -text
|
| 122 |
+
activations/qwen2.5-0.5b_english_wiki_750M_chinese_wikipedia_corpus/activation.zh.train.qwen-checkpoint-1962 filter=lfs diff=lfs merge=lfs -text
|
| 123 |
+
activations/qwen2.5-0.5b_english_wiki_750M_chinese_wikipedia_corpus/activation.zh.train.qwen-checkpoint-300 filter=lfs diff=lfs merge=lfs -text
|
| 124 |
+
activations/qwen2.5-0.5b_english_wiki_750M_chinese_wikipedia_corpus/activation.zh.train.qwen-checkpoint-600 filter=lfs diff=lfs merge=lfs -text
|
| 125 |
+
activations/qwen2.5-0.5b_english_wiki_750M_chinese_wikipedia_corpus/activation.zh.train.qwen-checkpoint-900 filter=lfs diff=lfs merge=lfs -text
|
| 126 |
+
activations/qwen2.5-0.5b_english_wiki_irish_corpus_custom_sampler/activation.en.train.qwen-checkpoint-1200 filter=lfs diff=lfs merge=lfs -text
|
| 127 |
+
activations/qwen2.5-0.5b_english_wiki_irish_corpus_custom_sampler/activation.en.train.qwen-checkpoint-1500 filter=lfs diff=lfs merge=lfs -text
|
| 128 |
+
activations/qwen2.5-0.5b_english_wiki_irish_corpus_custom_sampler/activation.en.train.qwen-checkpoint-1800 filter=lfs diff=lfs merge=lfs -text
|
| 129 |
+
activations/qwen2.5-0.5b_english_wiki_irish_corpus_custom_sampler/activation.en.train.qwen-checkpoint-2100 filter=lfs diff=lfs merge=lfs -text
|
| 130 |
+
activations/qwen2.5-0.5b_english_wiki_irish_corpus_custom_sampler/activation.en.train.qwen-checkpoint-2400 filter=lfs diff=lfs merge=lfs -text
|
| 131 |
+
activations/qwen2.5-0.5b_english_wiki_irish_corpus_custom_sampler/activation.en.train.qwen-checkpoint-2700 filter=lfs diff=lfs merge=lfs -text
|
| 132 |
+
activations/qwen2.5-0.5b_english_wiki_irish_corpus_custom_sampler/activation.en.train.qwen-checkpoint-300 filter=lfs diff=lfs merge=lfs -text
|
| 133 |
+
activations/qwen2.5-0.5b_english_wiki_irish_corpus_custom_sampler/activation.en.train.qwen-checkpoint-3000 filter=lfs diff=lfs merge=lfs -text
|
| 134 |
+
activations/qwen2.5-0.5b_english_wiki_irish_corpus_custom_sampler/activation.en.train.qwen-checkpoint-3300 filter=lfs diff=lfs merge=lfs -text
|
| 135 |
+
activations/qwen2.5-0.5b_english_wiki_irish_corpus_custom_sampler/activation.en.train.qwen-checkpoint-3600 filter=lfs diff=lfs merge=lfs -text
|
| 136 |
+
activations/qwen2.5-0.5b_english_wiki_irish_corpus_custom_sampler/activation.en.train.qwen-checkpoint-3900 filter=lfs diff=lfs merge=lfs -text
|
| 137 |
+
activations/qwen2.5-0.5b_english_wiki_irish_corpus_custom_sampler/activation.en.train.qwen-checkpoint-4200 filter=lfs diff=lfs merge=lfs -text
|
| 138 |
+
activations/qwen2.5-0.5b_english_wiki_irish_corpus_custom_sampler/activation.en.train.qwen-checkpoint-4500 filter=lfs diff=lfs merge=lfs -text
|
| 139 |
+
activations/qwen2.5-0.5b_english_wiki_irish_corpus_custom_sampler/activation.en.train.qwen-checkpoint-4800 filter=lfs diff=lfs merge=lfs -text
|
| 140 |
+
activations/qwen2.5-0.5b_english_wiki_irish_corpus_custom_sampler/activation.en.train.qwen-checkpoint-5100 filter=lfs diff=lfs merge=lfs -text
|
| 141 |
+
activations/qwen2.5-0.5b_english_wiki_irish_corpus_custom_sampler/activation.en.train.qwen-checkpoint-5400 filter=lfs diff=lfs merge=lfs -text
|
| 142 |
+
activations/qwen2.5-0.5b_english_wiki_irish_corpus_custom_sampler/activation.en.train.qwen-checkpoint-5700 filter=lfs diff=lfs merge=lfs -text
|
| 143 |
+
activations/qwen2.5-0.5b_english_wiki_irish_corpus_custom_sampler/activation.en.train.qwen-checkpoint-5784 filter=lfs diff=lfs merge=lfs -text
|
| 144 |
+
activations/qwen2.5-0.5b_english_wiki_irish_corpus_custom_sampler/activation.en.train.qwen-checkpoint-600 filter=lfs diff=lfs merge=lfs -text
|
| 145 |
+
activations/qwen2.5-0.5b_english_wiki_irish_corpus_custom_sampler/activation.en.train.qwen-checkpoint-900 filter=lfs diff=lfs merge=lfs -text
|
| 146 |
+
activations/qwen2.5-0.5b_english_wiki_irish_corpus_custom_sampler/activation.ga.train.qwen-checkpoint-1200 filter=lfs diff=lfs merge=lfs -text
|
| 147 |
+
activations/qwen2.5-0.5b_english_wiki_irish_corpus_custom_sampler/activation.ga.train.qwen-checkpoint-1500 filter=lfs diff=lfs merge=lfs -text
|
| 148 |
+
activations/qwen2.5-0.5b_english_wiki_irish_corpus_custom_sampler/activation.ga.train.qwen-checkpoint-1800 filter=lfs diff=lfs merge=lfs -text
|
| 149 |
+
activations/qwen2.5-0.5b_english_wiki_irish_corpus_custom_sampler/activation.ga.train.qwen-checkpoint-2100 filter=lfs diff=lfs merge=lfs -text
|
| 150 |
+
activations/qwen2.5-0.5b_english_wiki_irish_corpus_custom_sampler/activation.ga.train.qwen-checkpoint-2400 filter=lfs diff=lfs merge=lfs -text
|
| 151 |
+
activations/qwen2.5-0.5b_english_wiki_irish_corpus_custom_sampler/activation.ga.train.qwen-checkpoint-2700 filter=lfs diff=lfs merge=lfs -text
|
| 152 |
+
activations/qwen2.5-0.5b_english_wiki_irish_corpus_custom_sampler/activation.ga.train.qwen-checkpoint-300 filter=lfs diff=lfs merge=lfs -text
|
| 153 |
+
activations/qwen2.5-0.5b_english_wiki_irish_corpus_custom_sampler/activation.ga.train.qwen-checkpoint-3000 filter=lfs diff=lfs merge=lfs -text
|
| 154 |
+
activations/qwen2.5-0.5b_english_wiki_irish_corpus_custom_sampler/activation.ga.train.qwen-checkpoint-3300 filter=lfs diff=lfs merge=lfs -text
|
| 155 |
+
activations/qwen2.5-0.5b_english_wiki_irish_corpus_custom_sampler/activation.ga.train.qwen-checkpoint-3600 filter=lfs diff=lfs merge=lfs -text
|
| 156 |
+
activations/qwen2.5-0.5b_english_wiki_irish_corpus_custom_sampler/activation.ga.train.qwen-checkpoint-3900 filter=lfs diff=lfs merge=lfs -text
|
| 157 |
+
activations/qwen2.5-0.5b_english_wiki_irish_corpus_custom_sampler/activation.ga.train.qwen-checkpoint-4200 filter=lfs diff=lfs merge=lfs -text
|
| 158 |
+
activations/qwen2.5-0.5b_english_wiki_irish_corpus_custom_sampler/activation.ga.train.qwen-checkpoint-4500 filter=lfs diff=lfs merge=lfs -text
|
| 159 |
+
activations/qwen2.5-0.5b_english_wiki_irish_corpus_custom_sampler/activation.ga.train.qwen-checkpoint-4800 filter=lfs diff=lfs merge=lfs -text
|
| 160 |
+
activations/qwen2.5-0.5b_english_wiki_irish_corpus_custom_sampler/activation.ga.train.qwen-checkpoint-5100 filter=lfs diff=lfs merge=lfs -text
|
| 161 |
+
activations/qwen2.5-0.5b_english_wiki_irish_corpus_custom_sampler/activation.ga.train.qwen-checkpoint-5400 filter=lfs diff=lfs merge=lfs -text
|
| 162 |
+
activations/qwen2.5-0.5b_english_wiki_irish_corpus_custom_sampler/activation.ga.train.qwen-checkpoint-5700 filter=lfs diff=lfs merge=lfs -text
|
| 163 |
+
activations/qwen2.5-0.5b_english_wiki_irish_corpus_custom_sampler/activation.ga.train.qwen-checkpoint-5784 filter=lfs diff=lfs merge=lfs -text
|
| 164 |
+
activations/qwen2.5-0.5b_english_wiki_irish_corpus_custom_sampler/activation.ga.train.qwen-checkpoint-600 filter=lfs diff=lfs merge=lfs -text
|
| 165 |
+
activations/qwen2.5-0.5b_english_wiki_irish_corpus_custom_sampler/activation.ga.train.qwen-checkpoint-900 filter=lfs diff=lfs merge=lfs -text
|
| 166 |
+
activations-old/qwen2.5-0.5b_english_wiki_1.5B_basque_corpus/activation.en.train.qwen-checkpoint-1200 filter=lfs diff=lfs merge=lfs -text
|
| 167 |
+
activations-old/qwen2.5-0.5b_english_wiki_1.5B_basque_corpus/activation.en.train.qwen-checkpoint-1500 filter=lfs diff=lfs merge=lfs -text
|
| 168 |
+
activations-old/qwen2.5-0.5b_english_wiki_1.5B_basque_corpus/activation.en.train.qwen-checkpoint-1800 filter=lfs diff=lfs merge=lfs -text
|
| 169 |
+
activations-old/qwen2.5-0.5b_english_wiki_1.5B_basque_corpus/activation.en.train.qwen-checkpoint-2100 filter=lfs diff=lfs merge=lfs -text
|
| 170 |
+
activations-old/qwen2.5-0.5b_english_wiki_1.5B_basque_corpus/activation.en.train.qwen-checkpoint-2400 filter=lfs diff=lfs merge=lfs -text
|
| 171 |
+
activations-old/qwen2.5-0.5b_english_wiki_1.5B_basque_corpus/activation.en.train.qwen-checkpoint-2700 filter=lfs diff=lfs merge=lfs -text
|
| 172 |
+
activations-old/qwen2.5-0.5b_english_wiki_1.5B_basque_corpus/activation.en.train.qwen-checkpoint-300 filter=lfs diff=lfs merge=lfs -text
|
| 173 |
+
activations-old/qwen2.5-0.5b_english_wiki_1.5B_basque_corpus/activation.en.train.qwen-checkpoint-3000 filter=lfs diff=lfs merge=lfs -text
|
| 174 |
+
activations-old/qwen2.5-0.5b_english_wiki_1.5B_basque_corpus/activation.en.train.qwen-checkpoint-3231 filter=lfs diff=lfs merge=lfs -text
|
| 175 |
+
activations-old/qwen2.5-0.5b_english_wiki_1.5B_basque_corpus/activation.en.train.qwen-checkpoint-600 filter=lfs diff=lfs merge=lfs -text
|
| 176 |
+
activations-old/qwen2.5-0.5b_english_wiki_1.5B_basque_corpus/activation.en.train.qwen-checkpoint-900 filter=lfs diff=lfs merge=lfs -text
|
| 177 |
+
activations-old/qwen2.5-0.5b_english_wiki_1.5B_basque_corpus/activation.eu.train.qwen-checkpoint-1200 filter=lfs diff=lfs merge=lfs -text
|
| 178 |
+
activations-old/qwen2.5-0.5b_english_wiki_1.5B_basque_corpus/activation.eu.train.qwen-checkpoint-1500 filter=lfs diff=lfs merge=lfs -text
|
| 179 |
+
activations-old/qwen2.5-0.5b_english_wiki_1.5B_basque_corpus/activation.eu.train.qwen-checkpoint-1800 filter=lfs diff=lfs merge=lfs -text
|
| 180 |
+
activations-old/qwen2.5-0.5b_english_wiki_1.5B_basque_corpus/activation.eu.train.qwen-checkpoint-2100 filter=lfs diff=lfs merge=lfs -text
|
| 181 |
+
activations-old/qwen2.5-0.5b_english_wiki_1.5B_basque_corpus/activation.eu.train.qwen-checkpoint-2400 filter=lfs diff=lfs merge=lfs -text
|
| 182 |
+
activations-old/qwen2.5-0.5b_english_wiki_1.5B_basque_corpus/activation.eu.train.qwen-checkpoint-2700 filter=lfs diff=lfs merge=lfs -text
|
| 183 |
+
activations-old/qwen2.5-0.5b_english_wiki_1.5B_basque_corpus/activation.eu.train.qwen-checkpoint-300 filter=lfs diff=lfs merge=lfs -text
|
| 184 |
+
activations-old/qwen2.5-0.5b_english_wiki_1.5B_basque_corpus/activation.eu.train.qwen-checkpoint-3000 filter=lfs diff=lfs merge=lfs -text
|
| 185 |
+
activations-old/qwen2.5-0.5b_english_wiki_1.5B_basque_corpus/activation.eu.train.qwen-checkpoint-3231 filter=lfs diff=lfs merge=lfs -text
|
| 186 |
+
activations-old/qwen2.5-0.5b_english_wiki_1.5B_basque_corpus/activation.eu.train.qwen-checkpoint-600 filter=lfs diff=lfs merge=lfs -text
|
| 187 |
+
activations-old/qwen2.5-0.5b_english_wiki_1.5B_basque_corpus/activation.eu.train.qwen-checkpoint-900 filter=lfs diff=lfs merge=lfs -text
|
| 188 |
+
activations-old/qwen2.5-0.5b_english_wiki_300M_en_1.5Bbasque_corpus/activation.en.train.qwen-checkpoint-1200 filter=lfs diff=lfs merge=lfs -text
|
| 189 |
+
activations-old/qwen2.5-0.5b_english_wiki_300M_en_1.5Bbasque_corpus/activation.en.train.qwen-checkpoint-1500 filter=lfs diff=lfs merge=lfs -text
|
| 190 |
+
activations-old/qwen2.5-0.5b_english_wiki_300M_en_1.5Bbasque_corpus/activation.en.train.qwen-checkpoint-1800 filter=lfs diff=lfs merge=lfs -text
|
| 191 |
+
activations-old/qwen2.5-0.5b_english_wiki_300M_en_1.5Bbasque_corpus/activation.en.train.qwen-checkpoint-2100 filter=lfs diff=lfs merge=lfs -text
|
| 192 |
+
activations-old/qwen2.5-0.5b_english_wiki_300M_en_1.5Bbasque_corpus/activation.en.train.qwen-checkpoint-2400 filter=lfs diff=lfs merge=lfs -text
|
| 193 |
+
activations-old/qwen2.5-0.5b_english_wiki_300M_en_1.5Bbasque_corpus/activation.en.train.qwen-checkpoint-2700 filter=lfs diff=lfs merge=lfs -text
|
| 194 |
+
activations-old/qwen2.5-0.5b_english_wiki_300M_en_1.5Bbasque_corpus/activation.en.train.qwen-checkpoint-2972 filter=lfs diff=lfs merge=lfs -text
|
| 195 |
+
activations-old/qwen2.5-0.5b_english_wiki_300M_en_1.5Bbasque_corpus/activation.en.train.qwen-checkpoint-300 filter=lfs diff=lfs merge=lfs -text
|
| 196 |
+
activations-old/qwen2.5-0.5b_english_wiki_300M_en_1.5Bbasque_corpus/activation.en.train.qwen-checkpoint-600 filter=lfs diff=lfs merge=lfs -text
|
| 197 |
+
activations-old/qwen2.5-0.5b_english_wiki_300M_en_1.5Bbasque_corpus/activation.en.train.qwen-checkpoint-900 filter=lfs diff=lfs merge=lfs -text
|
| 198 |
+
activations-old/qwen2.5-0.5b_english_wiki_300M_en_1.5Bbasque_corpus/activation.eu.train.qwen-checkpoint-1200 filter=lfs diff=lfs merge=lfs -text
|
| 199 |
+
activations-old/qwen2.5-0.5b_english_wiki_300M_en_1.5Bbasque_corpus/activation.eu.train.qwen-checkpoint-1500 filter=lfs diff=lfs merge=lfs -text
|
| 200 |
+
activations-old/qwen2.5-0.5b_english_wiki_300M_en_1.5Bbasque_corpus/activation.eu.train.qwen-checkpoint-1800 filter=lfs diff=lfs merge=lfs -text
|
| 201 |
+
activations-old/qwen2.5-0.5b_english_wiki_300M_en_1.5Bbasque_corpus/activation.eu.train.qwen-checkpoint-2100 filter=lfs diff=lfs merge=lfs -text
|
| 202 |
+
activations-old/qwen2.5-0.5b_english_wiki_300M_en_1.5Bbasque_corpus/activation.eu.train.qwen-checkpoint-2400 filter=lfs diff=lfs merge=lfs -text
|
| 203 |
+
activations-old/qwen2.5-0.5b_english_wiki_300M_en_1.5Bbasque_corpus/activation.eu.train.qwen-checkpoint-2700 filter=lfs diff=lfs merge=lfs -text
|
| 204 |
+
activations-old/qwen2.5-0.5b_english_wiki_300M_en_1.5Bbasque_corpus/activation.eu.train.qwen-checkpoint-2972 filter=lfs diff=lfs merge=lfs -text
|
| 205 |
+
activations-old/qwen2.5-0.5b_english_wiki_300M_en_1.5Bbasque_corpus/activation.eu.train.qwen-checkpoint-300 filter=lfs diff=lfs merge=lfs -text
|
| 206 |
+
activations-old/qwen2.5-0.5b_english_wiki_300M_en_1.5Bbasque_corpus/activation.eu.train.qwen-checkpoint-600 filter=lfs diff=lfs merge=lfs -text
|
| 207 |
+
activations-old/qwen2.5-0.5b_english_wiki_300M_en_1.5Bbasque_corpus/activation.eu.train.qwen-checkpoint-900 filter=lfs diff=lfs merge=lfs -text
|
| 208 |
+
activations-old/qwen2.5-0.5b_english_wiki_300M_en_1.5Birish_corpus/activation.en.train.qwen-checkpoint-1200 filter=lfs diff=lfs merge=lfs -text
|
| 209 |
+
activations-old/qwen2.5-0.5b_english_wiki_300M_en_1.5Birish_corpus/activation.en.train.qwen-checkpoint-1500 filter=lfs diff=lfs merge=lfs -text
|
| 210 |
+
activations-old/qwen2.5-0.5b_english_wiki_300M_en_1.5Birish_corpus/activation.en.train.qwen-checkpoint-1800 filter=lfs diff=lfs merge=lfs -text
|
| 211 |
+
activations-old/qwen2.5-0.5b_english_wiki_300M_en_1.5Birish_corpus/activation.en.train.qwen-checkpoint-2100 filter=lfs diff=lfs merge=lfs -text
|
| 212 |
+
activations-old/qwen2.5-0.5b_english_wiki_300M_en_1.5Birish_corpus/activation.en.train.qwen-checkpoint-2400 filter=lfs diff=lfs merge=lfs -text
|
| 213 |
+
activations-old/qwen2.5-0.5b_english_wiki_300M_en_1.5Birish_corpus/activation.en.train.qwen-checkpoint-2700 filter=lfs diff=lfs merge=lfs -text
|
| 214 |
+
activations-old/qwen2.5-0.5b_english_wiki_300M_en_1.5Birish_corpus/activation.en.train.qwen-checkpoint-300 filter=lfs diff=lfs merge=lfs -text
|
| 215 |
+
activations-old/qwen2.5-0.5b_english_wiki_300M_en_1.5Birish_corpus/activation.en.train.qwen-checkpoint-3000 filter=lfs diff=lfs merge=lfs -text
|
| 216 |
+
activations-old/qwen2.5-0.5b_english_wiki_300M_en_1.5Birish_corpus/activation.en.train.qwen-checkpoint-3300 filter=lfs diff=lfs merge=lfs -text
|
| 217 |
+
activations-old/qwen2.5-0.5b_english_wiki_300M_en_1.5Birish_corpus/activation.en.train.qwen-checkpoint-3600 filter=lfs diff=lfs merge=lfs -text
|
| 218 |
+
activations-old/qwen2.5-0.5b_english_wiki_300M_en_1.5Birish_corpus/activation.en.train.qwen-checkpoint-3900 filter=lfs diff=lfs merge=lfs -text
|
| 219 |
+
activations-old/qwen2.5-0.5b_english_wiki_300M_en_1.5Birish_corpus/activation.en.train.qwen-checkpoint-4200 filter=lfs diff=lfs merge=lfs -text
|
| 220 |
+
activations-old/qwen2.5-0.5b_english_wiki_300M_en_1.5Birish_corpus/activation.en.train.qwen-checkpoint-4500 filter=lfs diff=lfs merge=lfs -text
|
| 221 |
+
activations-old/qwen2.5-0.5b_english_wiki_300M_en_1.5Birish_corpus/activation.en.train.qwen-checkpoint-4800 filter=lfs diff=lfs merge=lfs -text
|
| 222 |
+
activations-old/qwen2.5-0.5b_english_wiki_300M_en_1.5Birish_corpus/activation.en.train.qwen-checkpoint-4888 filter=lfs diff=lfs merge=lfs -text
|
| 223 |
+
activations-old/qwen2.5-0.5b_english_wiki_300M_en_1.5Birish_corpus/activation.en.train.qwen-checkpoint-600 filter=lfs diff=lfs merge=lfs -text
|
| 224 |
+
activations-old/qwen2.5-0.5b_english_wiki_300M_en_1.5Birish_corpus/activation.en.train.qwen-checkpoint-900 filter=lfs diff=lfs merge=lfs -text
|
| 225 |
+
activations-old/qwen2.5-0.5b_english_wiki_300M_en_1.5Birish_corpus/activation.ga.train.qwen-checkpoint-1200 filter=lfs diff=lfs merge=lfs -text
|
| 226 |
+
activations-old/qwen2.5-0.5b_english_wiki_300M_en_1.5Birish_corpus/activation.ga.train.qwen-checkpoint-1500 filter=lfs diff=lfs merge=lfs -text
|
| 227 |
+
activations-old/qwen2.5-0.5b_english_wiki_300M_en_1.5Birish_corpus/activation.ga.train.qwen-checkpoint-1800 filter=lfs diff=lfs merge=lfs -text
|
| 228 |
+
activations-old/qwen2.5-0.5b_english_wiki_300M_en_1.5Birish_corpus/activation.ga.train.qwen-checkpoint-2100 filter=lfs diff=lfs merge=lfs -text
|
| 229 |
+
activations-old/qwen2.5-0.5b_english_wiki_300M_en_1.5Birish_corpus/activation.ga.train.qwen-checkpoint-2400 filter=lfs diff=lfs merge=lfs -text
|
| 230 |
+
activations-old/qwen2.5-0.5b_english_wiki_300M_en_1.5Birish_corpus/activation.ga.train.qwen-checkpoint-2700 filter=lfs diff=lfs merge=lfs -text
|
| 231 |
+
activations-old/qwen2.5-0.5b_english_wiki_300M_en_1.5Birish_corpus/activation.ga.train.qwen-checkpoint-300 filter=lfs diff=lfs merge=lfs -text
|
| 232 |
+
activations-old/qwen2.5-0.5b_english_wiki_300M_en_1.5Birish_corpus/activation.ga.train.qwen-checkpoint-3000 filter=lfs diff=lfs merge=lfs -text
|
| 233 |
+
activations-old/qwen2.5-0.5b_english_wiki_300M_en_1.5Birish_corpus/activation.ga.train.qwen-checkpoint-3300 filter=lfs diff=lfs merge=lfs -text
|
| 234 |
+
activations-old/qwen2.5-0.5b_english_wiki_300M_en_1.5Birish_corpus/activation.ga.train.qwen-checkpoint-3600 filter=lfs diff=lfs merge=lfs -text
|
| 235 |
+
activations-old/qwen2.5-0.5b_english_wiki_300M_en_1.5Birish_corpus/activation.ga.train.qwen-checkpoint-3900 filter=lfs diff=lfs merge=lfs -text
|
| 236 |
+
activations-old/qwen2.5-0.5b_english_wiki_300M_en_1.5Birish_corpus/activation.ga.train.qwen-checkpoint-4200 filter=lfs diff=lfs merge=lfs -text
|
| 237 |
+
activations-old/qwen2.5-0.5b_english_wiki_300M_en_1.5Birish_corpus/activation.ga.train.qwen-checkpoint-4500 filter=lfs diff=lfs merge=lfs -text
|
| 238 |
+
activations-old/qwen2.5-0.5b_english_wiki_300M_en_1.5Birish_corpus/activation.ga.train.qwen-checkpoint-4800 filter=lfs diff=lfs merge=lfs -text
|
| 239 |
+
activations-old/qwen2.5-0.5b_english_wiki_300M_en_1.5Birish_corpus/activation.ga.train.qwen-checkpoint-4888 filter=lfs diff=lfs merge=lfs -text
|
| 240 |
+
activations-old/qwen2.5-0.5b_english_wiki_300M_en_1.5Birish_corpus/activation.ga.train.qwen-checkpoint-600 filter=lfs diff=lfs merge=lfs -text
|
| 241 |
+
activations-old/qwen2.5-0.5b_english_wiki_300M_en_1.5Birish_corpus/activation.ga.train.qwen-checkpoint-900 filter=lfs diff=lfs merge=lfs -text
|
| 242 |
+
activations-old/qwen2.5-0.5b_english_wiki_750M_chinese_wikipedia_corpus/activation.en.train.qwen-checkpoint-1200 filter=lfs diff=lfs merge=lfs -text
|
| 243 |
+
activations-old/qwen2.5-0.5b_english_wiki_750M_chinese_wikipedia_corpus/activation.en.train.qwen-checkpoint-1500 filter=lfs diff=lfs merge=lfs -text
|
| 244 |
+
activations-old/qwen2.5-0.5b_english_wiki_750M_chinese_wikipedia_corpus/activation.en.train.qwen-checkpoint-1800 filter=lfs diff=lfs merge=lfs -text
|
| 245 |
+
activations-old/qwen2.5-0.5b_english_wiki_750M_chinese_wikipedia_corpus/activation.en.train.qwen-checkpoint-1962 filter=lfs diff=lfs merge=lfs -text
|
| 246 |
+
activations-old/qwen2.5-0.5b_english_wiki_750M_chinese_wikipedia_corpus/activation.en.train.qwen-checkpoint-300 filter=lfs diff=lfs merge=lfs -text
|
| 247 |
+
activations-old/qwen2.5-0.5b_english_wiki_750M_chinese_wikipedia_corpus/activation.en.train.qwen-checkpoint-600 filter=lfs diff=lfs merge=lfs -text
|
| 248 |
+
activations-old/qwen2.5-0.5b_english_wiki_750M_chinese_wikipedia_corpus/activation.en.train.qwen-checkpoint-900 filter=lfs diff=lfs merge=lfs -text
|
| 249 |
+
activations-old/qwen2.5-0.5b_english_wiki_750M_chinese_wikipedia_corpus/activation.zh.train.qwen-checkpoint-1200 filter=lfs diff=lfs merge=lfs -text
|
| 250 |
+
activations-old/qwen2.5-0.5b_english_wiki_750M_chinese_wikipedia_corpus/activation.zh.train.qwen-checkpoint-1500 filter=lfs diff=lfs merge=lfs -text
|
| 251 |
+
activations-old/qwen2.5-0.5b_english_wiki_750M_chinese_wikipedia_corpus/activation.zh.train.qwen-checkpoint-1800 filter=lfs diff=lfs merge=lfs -text
|
| 252 |
+
activations-old/qwen2.5-0.5b_english_wiki_750M_chinese_wikipedia_corpus/activation.zh.train.qwen-checkpoint-1962 filter=lfs diff=lfs merge=lfs -text
|
| 253 |
+
activations-old/qwen2.5-0.5b_english_wiki_750M_chinese_wikipedia_corpus/activation.zh.train.qwen-checkpoint-300 filter=lfs diff=lfs merge=lfs -text
|
| 254 |
+
activations-old/qwen2.5-0.5b_english_wiki_750M_chinese_wikipedia_corpus/activation.zh.train.qwen-checkpoint-600 filter=lfs diff=lfs merge=lfs -text
|
| 255 |
+
activations-old/qwen2.5-0.5b_english_wiki_750M_chinese_wikipedia_corpus/activation.zh.train.qwen-checkpoint-900 filter=lfs diff=lfs merge=lfs -text
|
| 256 |
+
activations-old/qwen2.5-0.5b_english_wiki_irish_corpus_custom_sampler/activation.en.train.qwen-checkpoint-1200 filter=lfs diff=lfs merge=lfs -text
|
| 257 |
+
activations-old/qwen2.5-0.5b_english_wiki_irish_corpus_custom_sampler/activation.en.train.qwen-checkpoint-1500 filter=lfs diff=lfs merge=lfs -text
|
| 258 |
+
activations-old/qwen2.5-0.5b_english_wiki_irish_corpus_custom_sampler/activation.en.train.qwen-checkpoint-1800 filter=lfs diff=lfs merge=lfs -text
|
| 259 |
+
activations-old/qwen2.5-0.5b_english_wiki_irish_corpus_custom_sampler/activation.en.train.qwen-checkpoint-2100 filter=lfs diff=lfs merge=lfs -text
|
| 260 |
+
activations-old/qwen2.5-0.5b_english_wiki_irish_corpus_custom_sampler/activation.en.train.qwen-checkpoint-2400 filter=lfs diff=lfs merge=lfs -text
|
| 261 |
+
activations-old/qwen2.5-0.5b_english_wiki_irish_corpus_custom_sampler/activation.en.train.qwen-checkpoint-2700 filter=lfs diff=lfs merge=lfs -text
|
| 262 |
+
activations-old/qwen2.5-0.5b_english_wiki_irish_corpus_custom_sampler/activation.en.train.qwen-checkpoint-300 filter=lfs diff=lfs merge=lfs -text
|
| 263 |
+
activations-old/qwen2.5-0.5b_english_wiki_irish_corpus_custom_sampler/activation.en.train.qwen-checkpoint-3000 filter=lfs diff=lfs merge=lfs -text
|
| 264 |
+
activations-old/qwen2.5-0.5b_english_wiki_irish_corpus_custom_sampler/activation.en.train.qwen-checkpoint-3300 filter=lfs diff=lfs merge=lfs -text
|
| 265 |
+
activations-old/qwen2.5-0.5b_english_wiki_irish_corpus_custom_sampler/activation.en.train.qwen-checkpoint-3600 filter=lfs diff=lfs merge=lfs -text
|
| 266 |
+
activations-old/qwen2.5-0.5b_english_wiki_irish_corpus_custom_sampler/activation.en.train.qwen-checkpoint-3900 filter=lfs diff=lfs merge=lfs -text
|
| 267 |
+
activations-old/qwen2.5-0.5b_english_wiki_irish_corpus_custom_sampler/activation.en.train.qwen-checkpoint-4200 filter=lfs diff=lfs merge=lfs -text
|
| 268 |
+
activations-old/qwen2.5-0.5b_english_wiki_irish_corpus_custom_sampler/activation.en.train.qwen-checkpoint-4500 filter=lfs diff=lfs merge=lfs -text
|
| 269 |
+
activations-old/qwen2.5-0.5b_english_wiki_irish_corpus_custom_sampler/activation.en.train.qwen-checkpoint-4800 filter=lfs diff=lfs merge=lfs -text
|
| 270 |
+
activations-old/qwen2.5-0.5b_english_wiki_irish_corpus_custom_sampler/activation.en.train.qwen-checkpoint-5100 filter=lfs diff=lfs merge=lfs -text
|
| 271 |
+
activations-old/qwen2.5-0.5b_english_wiki_irish_corpus_custom_sampler/activation.en.train.qwen-checkpoint-5400 filter=lfs diff=lfs merge=lfs -text
|
| 272 |
+
activations-old/qwen2.5-0.5b_english_wiki_irish_corpus_custom_sampler/activation.en.train.qwen-checkpoint-5700 filter=lfs diff=lfs merge=lfs -text
|
| 273 |
+
activations-old/qwen2.5-0.5b_english_wiki_irish_corpus_custom_sampler/activation.en.train.qwen-checkpoint-5784 filter=lfs diff=lfs merge=lfs -text
|
| 274 |
+
activations-old/qwen2.5-0.5b_english_wiki_irish_corpus_custom_sampler/activation.en.train.qwen-checkpoint-600 filter=lfs diff=lfs merge=lfs -text
|
| 275 |
+
activations-old/qwen2.5-0.5b_english_wiki_irish_corpus_custom_sampler/activation.en.train.qwen-checkpoint-900 filter=lfs diff=lfs merge=lfs -text
|
| 276 |
+
activations-old/qwen2.5-0.5b_english_wiki_irish_corpus_custom_sampler/activation.ga.train.qwen-checkpoint-1200 filter=lfs diff=lfs merge=lfs -text
|
| 277 |
+
activations-old/qwen2.5-0.5b_english_wiki_irish_corpus_custom_sampler/activation.ga.train.qwen-checkpoint-1500 filter=lfs diff=lfs merge=lfs -text
|
| 278 |
+
activations-old/qwen2.5-0.5b_english_wiki_irish_corpus_custom_sampler/activation.ga.train.qwen-checkpoint-1800 filter=lfs diff=lfs merge=lfs -text
|
| 279 |
+
activations-old/qwen2.5-0.5b_english_wiki_irish_corpus_custom_sampler/activation.ga.train.qwen-checkpoint-2100 filter=lfs diff=lfs merge=lfs -text
|
| 280 |
+
activations-old/qwen2.5-0.5b_english_wiki_irish_corpus_custom_sampler/activation.ga.train.qwen-checkpoint-2400 filter=lfs diff=lfs merge=lfs -text
|
| 281 |
+
activations-old/qwen2.5-0.5b_english_wiki_irish_corpus_custom_sampler/activation.ga.train.qwen-checkpoint-2700 filter=lfs diff=lfs merge=lfs -text
|
| 282 |
+
activations-old/qwen2.5-0.5b_english_wiki_irish_corpus_custom_sampler/activation.ga.train.qwen-checkpoint-300 filter=lfs diff=lfs merge=lfs -text
|
| 283 |
+
activations-old/qwen2.5-0.5b_english_wiki_irish_corpus_custom_sampler/activation.ga.train.qwen-checkpoint-3000 filter=lfs diff=lfs merge=lfs -text
|
| 284 |
+
activations-old/qwen2.5-0.5b_english_wiki_irish_corpus_custom_sampler/activation.ga.train.qwen-checkpoint-3300 filter=lfs diff=lfs merge=lfs -text
|
| 285 |
+
activations-old/qwen2.5-0.5b_english_wiki_irish_corpus_custom_sampler/activation.ga.train.qwen-checkpoint-3600 filter=lfs diff=lfs merge=lfs -text
|
| 286 |
+
activations-old/qwen2.5-0.5b_english_wiki_irish_corpus_custom_sampler/activation.ga.train.qwen-checkpoint-3900 filter=lfs diff=lfs merge=lfs -text
|
| 287 |
+
activations-old/qwen2.5-0.5b_english_wiki_irish_corpus_custom_sampler/activation.ga.train.qwen-checkpoint-4200 filter=lfs diff=lfs merge=lfs -text
|
| 288 |
+
activations-old/qwen2.5-0.5b_english_wiki_irish_corpus_custom_sampler/activation.ga.train.qwen-checkpoint-4500 filter=lfs diff=lfs merge=lfs -text
|
| 289 |
+
activations-old/qwen2.5-0.5b_english_wiki_irish_corpus_custom_sampler/activation.ga.train.qwen-checkpoint-4800 filter=lfs diff=lfs merge=lfs -text
|
| 290 |
+
activations-old/qwen2.5-0.5b_english_wiki_irish_corpus_custom_sampler/activation.ga.train.qwen-checkpoint-5100 filter=lfs diff=lfs merge=lfs -text
|
| 291 |
+
activations-old/qwen2.5-0.5b_english_wiki_irish_corpus_custom_sampler/activation.ga.train.qwen-checkpoint-5400 filter=lfs diff=lfs merge=lfs -text
|
| 292 |
+
activations-old/qwen2.5-0.5b_english_wiki_irish_corpus_custom_sampler/activation.ga.train.qwen-checkpoint-5700 filter=lfs diff=lfs merge=lfs -text
|
| 293 |
+
activations-old/qwen2.5-0.5b_english_wiki_irish_corpus_custom_sampler/activation.ga.train.qwen-checkpoint-5784 filter=lfs diff=lfs merge=lfs -text
|
| 294 |
+
activations-old/qwen2.5-0.5b_english_wiki_irish_corpus_custom_sampler/activation.ga.train.qwen-checkpoint-600 filter=lfs diff=lfs merge=lfs -text
|
| 295 |
+
activations-old/qwen2.5-0.5b_english_wiki_irish_corpus_custom_sampler/activation.ga.train.qwen-checkpoint-900 filter=lfs diff=lfs merge=lfs -text
|
| 296 |
+
activations-old/test/activation.en.train.qwen-checkpoint-300 filter=lfs diff=lfs merge=lfs -text
|
| 297 |
+
activations-old/test/activation.ga.train.qwen-checkpoint-300 filter=lfs diff=lfs merge=lfs -text
|
| 298 |
+
job_scripts/logs/1483-activation-multiple-all.out filter=lfs diff=lfs merge=lfs -text
|
| 299 |
+
job_scripts/logs/1484-activation-multiple-all.out filter=lfs diff=lfs merge=lfs -text
|
| 300 |
+
new_activations_pt_neurons_bar.png filter=lfs diff=lfs merge=lfs -text
|
| 301 |
+
old_ids/qwen2.5-0.5/id.en.train.qwen2.5-0.5 filter=lfs diff=lfs merge=lfs -text
|
| 302 |
+
old_ids/qwen2.5-0.5/id.eu.train.qwen2.5-0.5 filter=lfs diff=lfs merge=lfs -text
|
| 303 |
+
old_ids/qwen2.5-0.5/id.ga.train.qwen2.5-0.5 filter=lfs diff=lfs merge=lfs -text
|
| 304 |
+
old_ids/qwen2.5-0.5/id.zh.train.qwen2.5-0.5 filter=lfs diff=lfs merge=lfs -text
|
| 305 |
+
oscar_ids/l2-13b/id.en.train.l2-13b filter=lfs diff=lfs merge=lfs -text
|
| 306 |
+
oscar_ids/l2-13b/id.ga.train.l2-13b filter=lfs diff=lfs merge=lfs -text
|
| 307 |
+
oscar_ids/l2-7b/id.en.train.l2-7b filter=lfs diff=lfs merge=lfs -text
|
| 308 |
+
oscar_ids/l2-7b/id.eu.train.l2-7b filter=lfs diff=lfs merge=lfs -text
|
| 309 |
+
oscar_ids/q2.5/id.en.train.q2.5 filter=lfs diff=lfs merge=lfs -text
|
| 310 |
+
oscar_ids/q2.5/id.eu.train.q2.5 filter=lfs diff=lfs merge=lfs -text
|
| 311 |
+
oscar_ids/q2.5/id.ga.train.q2.5 filter=lfs diff=lfs merge=lfs -text
|
| 312 |
+
oscar_ids/q2.5/id.zh.train.q2.5 filter=lfs diff=lfs merge=lfs -text
|
| 313 |
+
qwen2.5-0.5b_english_wiki_750M_chinese_wikipedia_corpus_qwen-checkpoint-1200_neurons_bar.png filter=lfs diff=lfs merge=lfs -text
|
| 314 |
+
qwen2.5-0.5b_english_wiki_750M_chinese_wikipedia_corpus_qwen-checkpoint-1500_neurons_bar.png filter=lfs diff=lfs merge=lfs -text
|
| 315 |
+
qwen2.5-0.5b_english_wiki_750M_chinese_wikipedia_corpus_qwen-checkpoint-1800_neurons_bar.png filter=lfs diff=lfs merge=lfs -text
|
| 316 |
+
qwen2.5-0.5b_english_wiki_750M_chinese_wikipedia_corpus_qwen-checkpoint-1962_neurons_bar.png filter=lfs diff=lfs merge=lfs -text
|
| 317 |
+
qwen2.5-0.5b_english_wiki_750M_chinese_wikipedia_corpus_qwen-checkpoint-300_neurons_bar.png filter=lfs diff=lfs merge=lfs -text
|
| 318 |
+
qwen2.5-0.5b_english_wiki_750M_chinese_wikipedia_corpus_qwen-checkpoint-600_neurons_bar.png filter=lfs diff=lfs merge=lfs -text
|
| 319 |
+
qwen2.5-0.5b_english_wiki_750M_chinese_wikipedia_corpus_qwen-checkpoint-900_neurons_bar.png filter=lfs diff=lfs merge=lfs -text
|
| 320 |
+
results-old/qwen2.5-0.5b_english_wiki_1.5B_basque_corpus/qwen2.5-0.5b_english_wiki_1.5B_basque_corpus_qwen-checkpoint-1200_neurons_bar.png filter=lfs diff=lfs merge=lfs -text
|
| 321 |
+
results-old/qwen2.5-0.5b_english_wiki_1.5B_basque_corpus/qwen2.5-0.5b_english_wiki_1.5B_basque_corpus_qwen-checkpoint-1500_neurons_bar.png filter=lfs diff=lfs merge=lfs -text
|
| 322 |
+
results-old/qwen2.5-0.5b_english_wiki_1.5B_basque_corpus/qwen2.5-0.5b_english_wiki_1.5B_basque_corpus_qwen-checkpoint-1800_neurons_bar.png filter=lfs diff=lfs merge=lfs -text
|
| 323 |
+
results-old/qwen2.5-0.5b_english_wiki_1.5B_basque_corpus/qwen2.5-0.5b_english_wiki_1.5B_basque_corpus_qwen-checkpoint-2100_neurons_bar.png filter=lfs diff=lfs merge=lfs -text
|
| 324 |
+
results-old/qwen2.5-0.5b_english_wiki_1.5B_basque_corpus/qwen2.5-0.5b_english_wiki_1.5B_basque_corpus_qwen-checkpoint-2400_neurons_bar.png filter=lfs diff=lfs merge=lfs -text
|
| 325 |
+
results-old/qwen2.5-0.5b_english_wiki_1.5B_basque_corpus/qwen2.5-0.5b_english_wiki_1.5B_basque_corpus_qwen-checkpoint-2700_neurons_bar.png filter=lfs diff=lfs merge=lfs -text
|
| 326 |
+
results-old/qwen2.5-0.5b_english_wiki_1.5B_basque_corpus/qwen2.5-0.5b_english_wiki_1.5B_basque_corpus_qwen-checkpoint-3000_neurons_bar.png filter=lfs diff=lfs merge=lfs -text
|
| 327 |
+
results-old/qwen2.5-0.5b_english_wiki_1.5B_basque_corpus/qwen2.5-0.5b_english_wiki_1.5B_basque_corpus_qwen-checkpoint-300_neurons_bar.png filter=lfs diff=lfs merge=lfs -text
|
| 328 |
+
results-old/qwen2.5-0.5b_english_wiki_1.5B_basque_corpus/qwen2.5-0.5b_english_wiki_1.5B_basque_corpus_qwen-checkpoint-3231_neurons_bar.png filter=lfs diff=lfs merge=lfs -text
|
| 329 |
+
results-old/qwen2.5-0.5b_english_wiki_1.5B_basque_corpus/qwen2.5-0.5b_english_wiki_1.5B_basque_corpus_qwen-checkpoint-600_neurons_bar.png filter=lfs diff=lfs merge=lfs -text
|
| 330 |
+
results-old/qwen2.5-0.5b_english_wiki_1.5B_basque_corpus/qwen2.5-0.5b_english_wiki_1.5B_basque_corpus_qwen-checkpoint-900_neurons_bar.png filter=lfs diff=lfs merge=lfs -text
|
| 331 |
+
results-old/qwen2.5-0.5b_english_wiki_300M_en_1.5Bbasque_corpus/qwen2.5-0.5b_english_wiki_300M_en_1.5Bbasque_corpus_qwen-checkpoint-1200_neurons_bar.png filter=lfs diff=lfs merge=lfs -text
|
| 332 |
+
results-old/qwen2.5-0.5b_english_wiki_300M_en_1.5Bbasque_corpus/qwen2.5-0.5b_english_wiki_300M_en_1.5Bbasque_corpus_qwen-checkpoint-1500_neurons_bar.png filter=lfs diff=lfs merge=lfs -text
|
| 333 |
+
results-old/qwen2.5-0.5b_english_wiki_300M_en_1.5Bbasque_corpus/qwen2.5-0.5b_english_wiki_300M_en_1.5Bbasque_corpus_qwen-checkpoint-1800_neurons_bar.png filter=lfs diff=lfs merge=lfs -text
|
| 334 |
+
results-old/qwen2.5-0.5b_english_wiki_300M_en_1.5Bbasque_corpus/qwen2.5-0.5b_english_wiki_300M_en_1.5Bbasque_corpus_qwen-checkpoint-2100_neurons_bar.png filter=lfs diff=lfs merge=lfs -text
|
| 335 |
+
results-old/qwen2.5-0.5b_english_wiki_300M_en_1.5Bbasque_corpus/qwen2.5-0.5b_english_wiki_300M_en_1.5Bbasque_corpus_qwen-checkpoint-2400_neurons_bar.png filter=lfs diff=lfs merge=lfs -text
|
| 336 |
+
results-old/qwen2.5-0.5b_english_wiki_300M_en_1.5Bbasque_corpus/qwen2.5-0.5b_english_wiki_300M_en_1.5Bbasque_corpus_qwen-checkpoint-2700_neurons_bar.png filter=lfs diff=lfs merge=lfs -text
|
| 337 |
+
results-old/qwen2.5-0.5b_english_wiki_300M_en_1.5Bbasque_corpus/qwen2.5-0.5b_english_wiki_300M_en_1.5Bbasque_corpus_qwen-checkpoint-2972_neurons_bar.png filter=lfs diff=lfs merge=lfs -text
|
| 338 |
+
results-old/qwen2.5-0.5b_english_wiki_300M_en_1.5Bbasque_corpus/qwen2.5-0.5b_english_wiki_300M_en_1.5Bbasque_corpus_qwen-checkpoint-300_neurons_bar.png filter=lfs diff=lfs merge=lfs -text
|
| 339 |
+
results-old/qwen2.5-0.5b_english_wiki_300M_en_1.5Bbasque_corpus/qwen2.5-0.5b_english_wiki_300M_en_1.5Bbasque_corpus_qwen-checkpoint-600_neurons_bar.png filter=lfs diff=lfs merge=lfs -text
|
| 340 |
+
results-old/qwen2.5-0.5b_english_wiki_300M_en_1.5Bbasque_corpus/qwen2.5-0.5b_english_wiki_300M_en_1.5Bbasque_corpus_qwen-checkpoint-900_neurons_bar.png filter=lfs diff=lfs merge=lfs -text
|
| 341 |
+
results-old/qwen2.5-0.5b_english_wiki_300M_en_1.5Birish_corpus/qwen2.5-0.5b_english_wiki_300M_en_1.5Birish_corpus_qwen-checkpoint-1200_neurons_bar.png filter=lfs diff=lfs merge=lfs -text
|
| 342 |
+
results-old/qwen2.5-0.5b_english_wiki_300M_en_1.5Birish_corpus/qwen2.5-0.5b_english_wiki_300M_en_1.5Birish_corpus_qwen-checkpoint-1500_neurons_bar.png filter=lfs diff=lfs merge=lfs -text
|
| 343 |
+
results-old/qwen2.5-0.5b_english_wiki_300M_en_1.5Birish_corpus/qwen2.5-0.5b_english_wiki_300M_en_1.5Birish_corpus_qwen-checkpoint-1800_neurons_bar.png filter=lfs diff=lfs merge=lfs -text
|
| 344 |
+
results-old/qwen2.5-0.5b_english_wiki_300M_en_1.5Birish_corpus/qwen2.5-0.5b_english_wiki_300M_en_1.5Birish_corpus_qwen-checkpoint-2100_neurons_bar.png filter=lfs diff=lfs merge=lfs -text
|
| 345 |
+
results-old/qwen2.5-0.5b_english_wiki_300M_en_1.5Birish_corpus/qwen2.5-0.5b_english_wiki_300M_en_1.5Birish_corpus_qwen-checkpoint-2400_neurons_bar.png filter=lfs diff=lfs merge=lfs -text
|
| 346 |
+
results-old/qwen2.5-0.5b_english_wiki_300M_en_1.5Birish_corpus/qwen2.5-0.5b_english_wiki_300M_en_1.5Birish_corpus_qwen-checkpoint-2700_neurons_bar.png filter=lfs diff=lfs merge=lfs -text
|
| 347 |
+
results-old/qwen2.5-0.5b_english_wiki_300M_en_1.5Birish_corpus/qwen2.5-0.5b_english_wiki_300M_en_1.5Birish_corpus_qwen-checkpoint-3000_neurons_bar.png filter=lfs diff=lfs merge=lfs -text
|
| 348 |
+
results-old/qwen2.5-0.5b_english_wiki_300M_en_1.5Birish_corpus/qwen2.5-0.5b_english_wiki_300M_en_1.5Birish_corpus_qwen-checkpoint-300_neurons_bar.png filter=lfs diff=lfs merge=lfs -text
|
| 349 |
+
results-old/qwen2.5-0.5b_english_wiki_300M_en_1.5Birish_corpus/qwen2.5-0.5b_english_wiki_300M_en_1.5Birish_corpus_qwen-checkpoint-3300_neurons_bar.png filter=lfs diff=lfs merge=lfs -text
|
| 350 |
+
results-old/qwen2.5-0.5b_english_wiki_300M_en_1.5Birish_corpus/qwen2.5-0.5b_english_wiki_300M_en_1.5Birish_corpus_qwen-checkpoint-3600_neurons_bar.png filter=lfs diff=lfs merge=lfs -text
|
| 351 |
+
results-old/qwen2.5-0.5b_english_wiki_300M_en_1.5Birish_corpus/qwen2.5-0.5b_english_wiki_300M_en_1.5Birish_corpus_qwen-checkpoint-3900_neurons_bar.png filter=lfs diff=lfs merge=lfs -text
|
| 352 |
+
results-old/qwen2.5-0.5b_english_wiki_300M_en_1.5Birish_corpus/qwen2.5-0.5b_english_wiki_300M_en_1.5Birish_corpus_qwen-checkpoint-4200_neurons_bar.png filter=lfs diff=lfs merge=lfs -text
|
| 353 |
+
results-old/qwen2.5-0.5b_english_wiki_300M_en_1.5Birish_corpus/qwen2.5-0.5b_english_wiki_300M_en_1.5Birish_corpus_qwen-checkpoint-4500_neurons_bar.png filter=lfs diff=lfs merge=lfs -text
|
| 354 |
+
results-old/qwen2.5-0.5b_english_wiki_300M_en_1.5Birish_corpus/qwen2.5-0.5b_english_wiki_300M_en_1.5Birish_corpus_qwen-checkpoint-4800_neurons_bar.png filter=lfs diff=lfs merge=lfs -text
|
| 355 |
+
results-old/qwen2.5-0.5b_english_wiki_300M_en_1.5Birish_corpus/qwen2.5-0.5b_english_wiki_300M_en_1.5Birish_corpus_qwen-checkpoint-4888_neurons_bar.png filter=lfs diff=lfs merge=lfs -text
|
| 356 |
+
results-old/qwen2.5-0.5b_english_wiki_300M_en_1.5Birish_corpus/qwen2.5-0.5b_english_wiki_300M_en_1.5Birish_corpus_qwen-checkpoint-600_neurons_bar.png filter=lfs diff=lfs merge=lfs -text
|
| 357 |
+
results-old/qwen2.5-0.5b_english_wiki_300M_en_1.5Birish_corpus/qwen2.5-0.5b_english_wiki_300M_en_1.5Birish_corpus_qwen-checkpoint-900_neurons_bar.png filter=lfs diff=lfs merge=lfs -text
|
| 358 |
+
results-old/qwen2.5-0.5b_english_wiki_750M_chinese_wikipedia_corpus/qwen2.5-0.5b_english_wiki_750M_chinese_wikipedia_corpus_qwen-checkpoint-1200_neurons_bar.png filter=lfs diff=lfs merge=lfs -text
|
| 359 |
+
results-old/qwen2.5-0.5b_english_wiki_750M_chinese_wikipedia_corpus/qwen2.5-0.5b_english_wiki_750M_chinese_wikipedia_corpus_qwen-checkpoint-1500_neurons_bar.png filter=lfs diff=lfs merge=lfs -text
|
| 360 |
+
results-old/qwen2.5-0.5b_english_wiki_750M_chinese_wikipedia_corpus/qwen2.5-0.5b_english_wiki_750M_chinese_wikipedia_corpus_qwen-checkpoint-1800_neurons_bar.png filter=lfs diff=lfs merge=lfs -text
|
| 361 |
+
results-old/qwen2.5-0.5b_english_wiki_750M_chinese_wikipedia_corpus/qwen2.5-0.5b_english_wiki_750M_chinese_wikipedia_corpus_qwen-checkpoint-1962_neurons_bar.png filter=lfs diff=lfs merge=lfs -text
|
| 362 |
+
results-old/qwen2.5-0.5b_english_wiki_750M_chinese_wikipedia_corpus/qwen2.5-0.5b_english_wiki_750M_chinese_wikipedia_corpus_qwen-checkpoint-300_neurons_bar.png filter=lfs diff=lfs merge=lfs -text
|
| 363 |
+
results-old/qwen2.5-0.5b_english_wiki_750M_chinese_wikipedia_corpus/qwen2.5-0.5b_english_wiki_750M_chinese_wikipedia_corpus_qwen-checkpoint-600_neurons_bar.png filter=lfs diff=lfs merge=lfs -text
|
| 364 |
+
results-old/qwen2.5-0.5b_english_wiki_750M_chinese_wikipedia_corpus/qwen2.5-0.5b_english_wiki_750M_chinese_wikipedia_corpus_qwen-checkpoint-900_neurons_bar.png filter=lfs diff=lfs merge=lfs -text
|
| 365 |
+
results-old/qwen2.5-0.5b_english_wiki_irish_corpus_custom_sampler/qwen2.5-0.5b_english_wiki_irish_corpus_custom_sampler_qwen-checkpoint-1200_neurons_bar.png filter=lfs diff=lfs merge=lfs -text
|
| 366 |
+
results-old/qwen2.5-0.5b_english_wiki_irish_corpus_custom_sampler/qwen2.5-0.5b_english_wiki_irish_corpus_custom_sampler_qwen-checkpoint-1500_neurons_bar.png filter=lfs diff=lfs merge=lfs -text
|
| 367 |
+
results-old/qwen2.5-0.5b_english_wiki_irish_corpus_custom_sampler/qwen2.5-0.5b_english_wiki_irish_corpus_custom_sampler_qwen-checkpoint-1800_neurons_bar.png filter=lfs diff=lfs merge=lfs -text
|
| 368 |
+
results-old/qwen2.5-0.5b_english_wiki_irish_corpus_custom_sampler/qwen2.5-0.5b_english_wiki_irish_corpus_custom_sampler_qwen-checkpoint-2100_neurons_bar.png filter=lfs diff=lfs merge=lfs -text
|
| 369 |
+
results-old/qwen2.5-0.5b_english_wiki_irish_corpus_custom_sampler/qwen2.5-0.5b_english_wiki_irish_corpus_custom_sampler_qwen-checkpoint-2400_neurons_bar.png filter=lfs diff=lfs merge=lfs -text
|
| 370 |
+
results-old/qwen2.5-0.5b_english_wiki_irish_corpus_custom_sampler/qwen2.5-0.5b_english_wiki_irish_corpus_custom_sampler_qwen-checkpoint-2700_neurons_bar.png filter=lfs diff=lfs merge=lfs -text
|
| 371 |
+
results-old/qwen2.5-0.5b_english_wiki_irish_corpus_custom_sampler/qwen2.5-0.5b_english_wiki_irish_corpus_custom_sampler_qwen-checkpoint-3000_neurons_bar.png filter=lfs diff=lfs merge=lfs -text
|
| 372 |
+
results-old/qwen2.5-0.5b_english_wiki_irish_corpus_custom_sampler/qwen2.5-0.5b_english_wiki_irish_corpus_custom_sampler_qwen-checkpoint-300_neurons_bar.png filter=lfs diff=lfs merge=lfs -text
|
| 373 |
+
results-old/qwen2.5-0.5b_english_wiki_irish_corpus_custom_sampler/qwen2.5-0.5b_english_wiki_irish_corpus_custom_sampler_qwen-checkpoint-3300_neurons_bar.png filter=lfs diff=lfs merge=lfs -text
|
| 374 |
+
results-old/qwen2.5-0.5b_english_wiki_irish_corpus_custom_sampler/qwen2.5-0.5b_english_wiki_irish_corpus_custom_sampler_qwen-checkpoint-3600_neurons_bar.png filter=lfs diff=lfs merge=lfs -text
|
| 375 |
+
results-old/qwen2.5-0.5b_english_wiki_irish_corpus_custom_sampler/qwen2.5-0.5b_english_wiki_irish_corpus_custom_sampler_qwen-checkpoint-3900_neurons_bar.png filter=lfs diff=lfs merge=lfs -text
|
| 376 |
+
results-old/qwen2.5-0.5b_english_wiki_irish_corpus_custom_sampler/qwen2.5-0.5b_english_wiki_irish_corpus_custom_sampler_qwen-checkpoint-4200_neurons_bar.png filter=lfs diff=lfs merge=lfs -text
|
| 377 |
+
results-old/qwen2.5-0.5b_english_wiki_irish_corpus_custom_sampler/qwen2.5-0.5b_english_wiki_irish_corpus_custom_sampler_qwen-checkpoint-4500_neurons_bar.png filter=lfs diff=lfs merge=lfs -text
|
| 378 |
+
results-old/qwen2.5-0.5b_english_wiki_irish_corpus_custom_sampler/qwen2.5-0.5b_english_wiki_irish_corpus_custom_sampler_qwen-checkpoint-4800_neurons_bar.png filter=lfs diff=lfs merge=lfs -text
|
| 379 |
+
results-old/qwen2.5-0.5b_english_wiki_irish_corpus_custom_sampler/qwen2.5-0.5b_english_wiki_irish_corpus_custom_sampler_qwen-checkpoint-5100_neurons_bar.png filter=lfs diff=lfs merge=lfs -text
|
| 380 |
+
results-old/qwen2.5-0.5b_english_wiki_irish_corpus_custom_sampler/qwen2.5-0.5b_english_wiki_irish_corpus_custom_sampler_qwen-checkpoint-5400_neurons_bar.png filter=lfs diff=lfs merge=lfs -text
|
| 381 |
+
results-old/qwen2.5-0.5b_english_wiki_irish_corpus_custom_sampler/qwen2.5-0.5b_english_wiki_irish_corpus_custom_sampler_qwen-checkpoint-5700_neurons_bar.png filter=lfs diff=lfs merge=lfs -text
|
| 382 |
+
results-old/qwen2.5-0.5b_english_wiki_irish_corpus_custom_sampler/qwen2.5-0.5b_english_wiki_irish_corpus_custom_sampler_qwen-checkpoint-5784_neurons_bar.png filter=lfs diff=lfs merge=lfs -text
|
| 383 |
+
results-old/qwen2.5-0.5b_english_wiki_irish_corpus_custom_sampler/qwen2.5-0.5b_english_wiki_irish_corpus_custom_sampler_qwen-checkpoint-600_neurons_bar.png filter=lfs diff=lfs merge=lfs -text
|
| 384 |
+
results-old/qwen2.5-0.5b_english_wiki_irish_corpus_custom_sampler/qwen2.5-0.5b_english_wiki_irish_corpus_custom_sampler_qwen-checkpoint-900_neurons_bar.png filter=lfs diff=lfs merge=lfs -text
|
README.md
ADDED
|
@@ -0,0 +1,16 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
## Language Specific Neuron SLA
|
| 2 |
+
|
| 3 |
+
This is done specifically for the Qwen2.5 family of models
|
| 4 |
+
|
| 5 |
+
## Guide
|
| 6 |
+
|
| 7 |
+
1. Run `load_data.py` to fetch data from https://huggingface.co/datasets/wikimedia/wikipedia/viewer/20231101
|
| 8 |
+
2. Calculate the activation from the fetched data with `activation.py`
|
| 9 |
+
3. Identify language specific neurons with `identify.py`
|
| 10 |
+
|
| 11 |
+
## Ref
|
| 12 |
+
- https://github.com/ReML-AI/DCL-CoT
|
| 13 |
+
- https://github.com/RUCAIBox/Language-Specific-Neurons
|
| 14 |
+
|
| 15 |
+
## Note taking
|
| 16 |
+
python3 load_data_oscar.py --languages en,zh,eu,ga --model-id qwen2.5 --tokenizer Qwen/Qwen2.5-0.5B
|
activation.py
ADDED
|
@@ -0,0 +1,101 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
#!/usr/bin/env python3
|
| 2 |
+
"""
|
| 3 |
+
Track per-neuron activations in Qwen2 MLP layers using Hugging Face Transformers
|
| 4 |
+
with explicit device management.
|
| 5 |
+
"""
|
| 6 |
+
|
| 7 |
+
import argparse
|
| 8 |
+
import os
|
| 9 |
+
from types import MethodType
|
| 10 |
+
|
| 11 |
+
import torch
|
| 12 |
+
from torch import Tensor
|
| 13 |
+
from tqdm import tqdm
|
| 14 |
+
from transformers import AutoModelForCausalLM
|
| 15 |
+
|
| 16 |
+
# ---------------------- Activation Tracker ----------------------
|
| 17 |
+
class ActivationTracker:
|
| 18 |
+
def __init__(self, num_layers: int, intermediate_size: int):
|
| 19 |
+
# store on CPU to avoid memory issues
|
| 20 |
+
self.over_zero = torch.zeros(
|
| 21 |
+
num_layers, intermediate_size, dtype=torch.int32, device="cpu"
|
| 22 |
+
)
|
| 23 |
+
|
| 24 |
+
def make_qwen_hook(self, index: int):
|
| 25 |
+
over_zero = self.over_zero
|
| 26 |
+
|
| 27 |
+
def qwen_forward(self, x: Tensor):
|
| 28 |
+
gate_activation = self.act_fn(self.gate_proj(x))
|
| 29 |
+
with torch.no_grad():
|
| 30 |
+
over_zero[index, :] += (gate_activation > 0).sum(dim=(0, 1)).to("cpu")
|
| 31 |
+
return self.down_proj(gate_activation * self.up_proj(x))
|
| 32 |
+
|
| 33 |
+
return qwen_forward
|
| 34 |
+
|
| 35 |
+
# ---------------------- Arguments ----------------------
|
| 36 |
+
parser = argparse.ArgumentParser()
|
| 37 |
+
parser.add_argument("--model", type=str, required=True, help="HF model ID or local folder path")
|
| 38 |
+
parser.add_argument("--lang", type=str, required=True, help="Language code for dataset")
|
| 39 |
+
parser.add_argument("--data-path", type=str, required=True, help="Path to tokenized dataset (torch tensor)")
|
| 40 |
+
parser.add_argument("--output-dir", type=str, default="activations", help="Directory to save over_zero")
|
| 41 |
+
parser.add_argument("--batch-size", type=int, default=1, help="Batch size per device")
|
| 42 |
+
parser.add_argument("--chunk-size", type=int, default=4096, help="Max sequence length to process at once")
|
| 43 |
+
args = parser.parse_args()
|
| 44 |
+
|
| 45 |
+
# ---------------------- Setup Device ----------------------
|
| 46 |
+
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
|
| 47 |
+
print(f"Using device: {device}")
|
| 48 |
+
os.makedirs(args.output_dir, exist_ok=True)
|
| 49 |
+
|
| 50 |
+
# ---------------------- Load Data ----------------------
|
| 51 |
+
print("Loading data...")
|
| 52 |
+
ids = torch.load(args.data_path, map_location="cpu") # Load to CPU first
|
| 53 |
+
|
| 54 |
+
# ---------------------- Load Model ----------------------
|
| 55 |
+
print(f"Loading model: {args.model}")
|
| 56 |
+
model = AutoModelForCausalLM.from_pretrained(
|
| 57 |
+
args.model,
|
| 58 |
+
device_map="auto", # Let it automatically distribute across available GPUs
|
| 59 |
+
torch_dtype=torch.bfloat16 # reduce memory
|
| 60 |
+
)
|
| 61 |
+
model.eval()
|
| 62 |
+
|
| 63 |
+
num_layers = model.config.num_hidden_layers
|
| 64 |
+
intermediate_size = model.config.intermediate_size
|
| 65 |
+
max_len = model.config.max_position_embeddings
|
| 66 |
+
|
| 67 |
+
# Setup tracker
|
| 68 |
+
tracker = ActivationTracker(num_layers=num_layers, intermediate_size=intermediate_size)
|
| 69 |
+
|
| 70 |
+
# Monkey-patch MLP layers
|
| 71 |
+
for i, layer in enumerate(model.model.layers):
|
| 72 |
+
layer.mlp.forward = MethodType(tracker.make_qwen_hook(i), layer.mlp)
|
| 73 |
+
|
| 74 |
+
# Prepare input - use chunk_size instead of max_len for memory efficiency
|
| 75 |
+
chunk_size = min(args.chunk_size, max_len)
|
| 76 |
+
n = (ids.size(0) // chunk_size) * chunk_size
|
| 77 |
+
input_ids = ids[:n].reshape(-1, chunk_size)
|
| 78 |
+
|
| 79 |
+
print(f"Processing {input_ids.size(0)} sequences of length {chunk_size}")
|
| 80 |
+
|
| 81 |
+
# ---------------------- Run Inference ----------------------
|
| 82 |
+
with torch.no_grad():
|
| 83 |
+
for i in tqdm(range(0, input_ids.size(0), args.batch_size), desc="Processing", unit="batch"):
|
| 84 |
+
batch = input_ids[i:i + args.batch_size]
|
| 85 |
+
|
| 86 |
+
# Move batch to the same device as the model's first parameter
|
| 87 |
+
# This works with device_map="auto"
|
| 88 |
+
batch = batch.to(next(model.parameters()).device)
|
| 89 |
+
|
| 90 |
+
# Clear cache before each batch to prevent memory buildup
|
| 91 |
+
if torch.cuda.is_available():
|
| 92 |
+
torch.cuda.empty_cache()
|
| 93 |
+
|
| 94 |
+
model(input_ids=batch)
|
| 95 |
+
|
| 96 |
+
# ---------------------- Save activations ----------------------
|
| 97 |
+
model_name = os.path.basename(args.model.rstrip("/"))
|
| 98 |
+
out_path = os.path.join(args.output_dir, f"activation_{model_name}_{args.lang}.pt")
|
| 99 |
+
torch.save(tracker.over_zero, out_path)
|
| 100 |
+
print(f"Saved activation counts to {out_path}")
|
| 101 |
+
print("Activation single job done")
|
activation_all.log
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|
activation_all.py
ADDED
|
@@ -0,0 +1,192 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
#!/usr/bin/env python3
|
| 2 |
+
"""
|
| 3 |
+
Per-neuron activation tracker for LLaMA-2 and Qwen MLP layers.
|
| 4 |
+
Runs on a fixed set of models and multiple input ID files per model.
|
| 5 |
+
"""
|
| 6 |
+
|
| 7 |
+
import torch
|
| 8 |
+
import os
|
| 9 |
+
from types import MethodType
|
| 10 |
+
from vllm import LLM, SamplingParams # Keep original import since hook logic depends on vLLM
|
| 11 |
+
|
| 12 |
+
# ---------------------- Config ----------------------
|
| 13 |
+
BASE_PATH = "/home/khanh/sla/sla_cpt"
|
| 14 |
+
ID_BASE_PATH = "./oscar_ids"
|
| 15 |
+
|
| 16 |
+
RUN_CONFIGS = [
|
| 17 |
+
# {
|
| 18 |
+
# 'name': 'l2-13b',
|
| 19 |
+
# 'model': f'{BASE_PATH}/uccix/checkpoint-4280',
|
| 20 |
+
# 'ids_list': [
|
| 21 |
+
# {"path": './ids/l2-13b/id.ga.train.l2-13b', "lang": "ga"},
|
| 22 |
+
# {"path": './ids/l2-13b/id.en.train.l2-13b', "lang": "en"}
|
| 23 |
+
# ],
|
| 24 |
+
# 'type': 'llama'
|
| 25 |
+
# },
|
| 26 |
+
# {
|
| 27 |
+
# 'name': 'l2-7b',
|
| 28 |
+
# 'model': f'{BASE_PATH}/llama2_7b_full_basque_corpus_grad_clip_1/checkpoint-10200',
|
| 29 |
+
# 'ids_list': [
|
| 30 |
+
# {"path": './ids/l2-7b/id.eu.train.l2-7b', "lang": "eu"},
|
| 31 |
+
# {"path": './ids/l2-7b/id.en.train.l2-7b', "lang": "en"}
|
| 32 |
+
# ],
|
| 33 |
+
# 'type': 'llama'
|
| 34 |
+
# },
|
| 35 |
+
{
|
| 36 |
+
'name': 'q2.5-zh',
|
| 37 |
+
'model': f'{BASE_PATH}/qwen2.5-0.5b_english_wiki_750M_chinese_wikipedia_corpus_2e_240925/checkpoint-2944',
|
| 38 |
+
'ids_list': [
|
| 39 |
+
{"path": f'{ID_BASE_PATH}/q2.5/id.zh.train.qwen2.5-0.5', "lang": "zh"},
|
| 40 |
+
{"path": f'{ID_BASE_PATH}/q2.5/id.en.train.qwen2.5-0.5', "lang": "en"}
|
| 41 |
+
],
|
| 42 |
+
'type': 'qwen'
|
| 43 |
+
},
|
| 44 |
+
# {
|
| 45 |
+
# 'name': 'q2.5-en+zh',
|
| 46 |
+
# 'model': f'{BASE_PATH}/qwen2.5-0.5b_english_wiki_150M_en_750M_chinese_wikipedia_corpus_2e_240925/checkpoint-3494',
|
| 47 |
+
# 'ids_list': [
|
| 48 |
+
# {"path": '{ID_BASE_PATH}/q2.5/id.zh.train.qwen2.5-0.5', "lang": "zh"},
|
| 49 |
+
# {"path": '{ID_BASE_PATH}/q2.5/id.en.train.qwen2.5-0.5', "lang": "en"}
|
| 50 |
+
# ],
|
| 51 |
+
# 'type': 'qwen'
|
| 52 |
+
# },
|
| 53 |
+
# {
|
| 54 |
+
# 'name': 'q2.5-ga',
|
| 55 |
+
# 'model': f'{BASE_PATH}/qwen2.5-0.5b_english_wiki_1.5B_irish_corpus_240925/checkpoint-2854',
|
| 56 |
+
# 'ids_list': [
|
| 57 |
+
# {"path": '{ID_BASE_PATH}/q2.5/id.en.train.qwen2.5-0.5', "lang": "en"},
|
| 58 |
+
# {"path": '{ID_BASE_PATH}/q2.5/id.ga.train.qwen2.5-0.5', "lang": "ga"}
|
| 59 |
+
# ],
|
| 60 |
+
# 'type': 'qwen'
|
| 61 |
+
# },
|
| 62 |
+
# # {
|
| 63 |
+
# # 'name': 'q2.5-en+ga',
|
| 64 |
+
# # 'model': f'{BASE_PATH}/qwen2.5-0.5_full_english_corpus_grad_clip_1/checkpoint-3231',
|
| 65 |
+
# # 'ids_list': [
|
| 66 |
+
# # {"path": './ids/qwen2.5-0.5/id.en.train.qwen2.5-0.5', "lang": "en"},
|
| 67 |
+
# # {"path": './ids/qwen2.5-0.5/id.ga.train.qwen2.5-0.5', "lang": "ga"}
|
| 68 |
+
# # ],
|
| 69 |
+
# # 'type': 'qwen'
|
| 70 |
+
# # },
|
| 71 |
+
# {
|
| 72 |
+
# 'name': 'q2.5-eu',
|
| 73 |
+
# 'model': f'{BASE_PATH}/qwen2.5-0.5b_english_wiki_1.5Bbasque_corpus_240925/checkpoint-2424',
|
| 74 |
+
# 'ids_list': [
|
| 75 |
+
# {"path": '{ID_BASE_PATH}/q2.5/id.eu.train.qwen2.5-0.5', "lang": "eu"},
|
| 76 |
+
# {"path": '{ID_BASE_PATH}/q2.5/id.en.train.qwen2.5-0.5', "lang": "en"}
|
| 77 |
+
# ],
|
| 78 |
+
# 'type': 'qwen'
|
| 79 |
+
# },
|
| 80 |
+
# {
|
| 81 |
+
# 'name': 'q2.5-en+eu',
|
| 82 |
+
# 'model': f'{BASE_PATH}/qwen2.5-0.5_full_basque_corpus_grad_clip_1/checkpoint-7800',
|
| 83 |
+
# 'ids_list': [
|
| 84 |
+
# {"path": './ids/qwen2.5-0.5/id.eu.train.qwen2.5-0.5', "lang": "eu"},
|
| 85 |
+
# {"path": './ids/qwen2.5-0.5/id.en.train.qwen2.5-0.5', "lang": "en"}
|
| 86 |
+
# ],
|
| 87 |
+
# }
|
| 88 |
+
]
|
| 89 |
+
|
| 90 |
+
SAVE_FOLDER = "new_activations"
|
| 91 |
+
os.makedirs(SAVE_FOLDER, exist_ok=True)
|
| 92 |
+
|
| 93 |
+
# ---------------------- Hook Functions ----------------------
|
| 94 |
+
def make_llama_hook(idx):
|
| 95 |
+
def llama_forward(self, x):
|
| 96 |
+
gate_up, _ = self.gate_up_proj(x) # l, 2i
|
| 97 |
+
i = gate_up.size(-1)
|
| 98 |
+
gate_up[:, : i // 2] = torch.nn.SiLU()(gate_up[:, : i // 2])
|
| 99 |
+
activation = gate_up[:, : i // 2].float() # l, i
|
| 100 |
+
over_zero[idx, :] += (activation > 0).sum(dim=0)
|
| 101 |
+
x = gate_up[:, : i // 2] * gate_up[:, i // 2 :]
|
| 102 |
+
x, _ = self.down_proj(x)
|
| 103 |
+
return x
|
| 104 |
+
return llama_forward
|
| 105 |
+
|
| 106 |
+
def make_qwen_hook(idx):
|
| 107 |
+
def qwen_forward(self, x):
|
| 108 |
+
gate_up, _ = self.gate_up_proj(x) # (s, 2h)
|
| 109 |
+
intermediate_size = gate_up.size(-1) // 2
|
| 110 |
+
gate = gate_up[..., :intermediate_size] # (s, h)
|
| 111 |
+
up = gate_up[..., intermediate_size:] # (s, h)
|
| 112 |
+
gate_activation = torch.nn.functional.silu(gate)
|
| 113 |
+
over_zero[idx, :] += (gate_activation > 0).sum(dim=0)
|
| 114 |
+
x, _ = self.down_proj(gate_activation * up)
|
| 115 |
+
return x
|
| 116 |
+
return qwen_forward
|
| 117 |
+
|
| 118 |
+
# ---------------------- Run All Configs ----------------------
|
| 119 |
+
for config in RUN_CONFIGS:
|
| 120 |
+
model_name = config['model']
|
| 121 |
+
save_name = config.get('name', model_name)
|
| 122 |
+
model_type = config.get('type', 'llama')
|
| 123 |
+
ids_list = config.get('ids_list', [])
|
| 124 |
+
|
| 125 |
+
print(f"\n=== Processing model: {model_name}, type: {model_type} ===")
|
| 126 |
+
|
| 127 |
+
# Load model
|
| 128 |
+
model = LLM(
|
| 129 |
+
model=model_name,
|
| 130 |
+
tensor_parallel_size=1,
|
| 131 |
+
enforce_eager=True,
|
| 132 |
+
trust_remote_code=True
|
| 133 |
+
)
|
| 134 |
+
|
| 135 |
+
max_length = model.llm_engine.model_config.max_model_len
|
| 136 |
+
num_layers = model.llm_engine.model_config.hf_config.num_hidden_layers
|
| 137 |
+
intermediate_size = model.llm_engine.model_config.hf_config.intermediate_size
|
| 138 |
+
|
| 139 |
+
print(f"Layers: {num_layers}, Intermediate size: {intermediate_size}, Max length: {max_length}")
|
| 140 |
+
|
| 141 |
+
# Setup activation tracker
|
| 142 |
+
over_zero = torch.zeros(num_layers, intermediate_size, dtype=torch.int32).to('cuda')
|
| 143 |
+
|
| 144 |
+
# Hook MLP layers
|
| 145 |
+
for i in range(num_layers):
|
| 146 |
+
mlp = model.llm_engine.model_executor.driver_worker.model_runner.model.model.layers[i].mlp
|
| 147 |
+
if model_type == 'llama':
|
| 148 |
+
mlp.forward = MethodType(make_llama_hook(i), mlp)
|
| 149 |
+
elif model_type == 'qwen':
|
| 150 |
+
mlp.forward = MethodType(make_qwen_hook(i), mlp)
|
| 151 |
+
else:
|
| 152 |
+
raise ValueError(f"Unknown model type: {model_type}")
|
| 153 |
+
|
| 154 |
+
# Iterate over all ID files
|
| 155 |
+
for id_dict in ids_list:
|
| 156 |
+
ids_path = id_dict['path']
|
| 157 |
+
lang = id_dict.get('lang', 'unknown') # Use lang in dict for output filename
|
| 158 |
+
|
| 159 |
+
print(f"\nLoading IDs from {ids_path} (lang: {lang})...")
|
| 160 |
+
ids = torch.load(ids_path)
|
| 161 |
+
print(f"ID shape: {ids.shape}")
|
| 162 |
+
|
| 163 |
+
l = ids.size(0)
|
| 164 |
+
l = min(l, 99999744) // max_length * max_length
|
| 165 |
+
input_ids = ids[:l].reshape(-1, max_length)
|
| 166 |
+
print(f"Processing {input_ids.size(0)} sequences of length {max_length}")
|
| 167 |
+
|
| 168 |
+
# Run inference
|
| 169 |
+
print("Running inference...")
|
| 170 |
+
_ = model.generate(
|
| 171 |
+
prompt_token_ids=input_ids.tolist(),
|
| 172 |
+
sampling_params=SamplingParams(max_tokens=1)
|
| 173 |
+
)
|
| 174 |
+
|
| 175 |
+
# Save results for this ID file
|
| 176 |
+
output_path = os.path.join(SAVE_FOLDER, f'activation.{lang}.train.{save_name}.pt')
|
| 177 |
+
torch.save({
|
| 178 |
+
'n': l,
|
| 179 |
+
'over_zero': over_zero.cpu(),
|
| 180 |
+
'num_layers': num_layers,
|
| 181 |
+
'intermediate_size': intermediate_size
|
| 182 |
+
}, output_path)
|
| 183 |
+
|
| 184 |
+
print(f"Saved activation counts to {output_path}")
|
| 185 |
+
print(f"Processed {l} tokens total")
|
| 186 |
+
|
| 187 |
+
print(f"\nActivation analysis complete for model: {save_name}!")
|
| 188 |
+
|
| 189 |
+
del model
|
| 190 |
+
torch.cuda.empty_cache()
|
| 191 |
+
import gc
|
| 192 |
+
gc.collect()
|
activation_all.sh
ADDED
|
@@ -0,0 +1,16 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
#!/bin/bash
|
| 2 |
+
#SBATCH --job-name=activation_all # Job name
|
| 3 |
+
#SBATCH --output=activation_all.log # Standard output log
|
| 4 |
+
#SBATCH --error=activation_all.log # Standard error log
|
| 5 |
+
#SBATCH --ntasks=1 # Number of tasks (processes)
|
| 6 |
+
#SBATCH --cpus-per-task=32 # Number of CPU cores
|
| 7 |
+
#SBATCH --mem=256G # Memory
|
| 8 |
+
#SBATCH --gres=gpu:1 # Number of GPUs
|
| 9 |
+
#SBATCH --partition=physical-gpu # Partition/queue name
|
| 10 |
+
|
| 11 |
+
# Your command here
|
| 12 |
+
|
| 13 |
+
export VLLM_USE_V1=0
|
| 14 |
+
export PYTHON_UNBUFFERED=1
|
| 15 |
+
|
| 16 |
+
srun python3 activation_all.py
|
activation_llama.py
ADDED
|
@@ -0,0 +1,101 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
#!/usr/bin/env python3
|
| 2 |
+
"""
|
| 3 |
+
Simple per-neuron activation tracker for LLaMA-2 MLP layers.
|
| 4 |
+
Runs on a fixed set of models and input IDs.
|
| 5 |
+
"""
|
| 6 |
+
|
| 7 |
+
import torch
|
| 8 |
+
import os
|
| 9 |
+
from types import MethodType
|
| 10 |
+
from vllm import LLM, SamplingParams # Keep your original import since hook logic depends on vLLM
|
| 11 |
+
|
| 12 |
+
# ---------------------- Config ----------------------
|
| 13 |
+
RUN_CONFIGS = [
|
| 14 |
+
{
|
| 15 |
+
'model': 'qwen2.5-0.5', # replace with LLaMA2 HF model if needed
|
| 16 |
+
'ids_path': '../ids/qwen2.5-0.5/id.en.train.qwen2.5-0.5',
|
| 17 |
+
'lang': 'en'
|
| 18 |
+
},
|
| 19 |
+
{
|
| 20 |
+
'model': 'qwen2.5-0.5',
|
| 21 |
+
'ids_path': '../ids/qwen2.5-0.5/id.de.train.qwen2.5-0.5',
|
| 22 |
+
'lang': 'de'
|
| 23 |
+
},
|
| 24 |
+
# Add more entries here
|
| 25 |
+
]
|
| 26 |
+
|
| 27 |
+
SAVE_FOLDER = "data"
|
| 28 |
+
os.makedirs(SAVE_FOLDER, exist_ok=True)
|
| 29 |
+
|
| 30 |
+
# ---------------------- Original Hook Function ----------------------
|
| 31 |
+
def make_llama_hook(idx):
|
| 32 |
+
def llama_forward(self, x):
|
| 33 |
+
gate_up, _ = self.gate_up_proj(x) # b, l, 2i
|
| 34 |
+
i = gate_up.size(-1)
|
| 35 |
+
gate_up[:, :, : i // 2] = torch.nn.SiLU()(gate_up[:, :, : i // 2])
|
| 36 |
+
activation = gate_up[:, :, : i // 2].float() # b, l, i
|
| 37 |
+
over_zero[idx, :] += (activation > 0).sum(dim=(0,1))
|
| 38 |
+
x = gate_up[:, :, : i // 2] * gate_up[:, :, i // 2 :]
|
| 39 |
+
x, _ = self.down_proj(x)
|
| 40 |
+
return x
|
| 41 |
+
return llama_forward
|
| 42 |
+
|
| 43 |
+
# ---------------------- Run All Configs ----------------------
|
| 44 |
+
for config in RUN_CONFIGS:
|
| 45 |
+
model_name = config['model']
|
| 46 |
+
lang = config['lang']
|
| 47 |
+
ids_path = config['ids_path']
|
| 48 |
+
|
| 49 |
+
print(f"\n=== Processing model: {model_name}, lang: {lang} ===")
|
| 50 |
+
|
| 51 |
+
# Load model
|
| 52 |
+
model = LLM(
|
| 53 |
+
model=model_name,
|
| 54 |
+
tensor_parallel_size=1,
|
| 55 |
+
enforce_eager=True,
|
| 56 |
+
trust_remote_code=True
|
| 57 |
+
)
|
| 58 |
+
|
| 59 |
+
max_length = model.llm_engine.model_config.max_model_len
|
| 60 |
+
num_layers = model.llm_engine.model_config.hf_config.num_hidden_layers
|
| 61 |
+
intermediate_size = model.llm_engine.model_config.hf_config.intermediate_size
|
| 62 |
+
|
| 63 |
+
print(f"Layers: {num_layers}, Intermediate size: {intermediate_size}, Max length: {max_length}")
|
| 64 |
+
|
| 65 |
+
# Setup activation tracker
|
| 66 |
+
over_zero = torch.zeros(num_layers, intermediate_size, dtype=torch.int32).to('cuda')
|
| 67 |
+
|
| 68 |
+
# Hook MLP layers
|
| 69 |
+
for i in range(num_layers):
|
| 70 |
+
mlp = model.llm_engine.model_executor.driver_worker.model_runner.model.model.layers[i].mlp
|
| 71 |
+
mlp.forward = MethodType(make_llama_hook(i), mlp)
|
| 72 |
+
|
| 73 |
+
# Load input IDs
|
| 74 |
+
print("Loading IDs...")
|
| 75 |
+
ids = torch.load(ids_path)
|
| 76 |
+
print(f"ID shape: {ids.shape}")
|
| 77 |
+
|
| 78 |
+
l = ids.size(0)
|
| 79 |
+
l = min(l, 99999744) // max_length * max_length
|
| 80 |
+
input_ids = ids[:l].reshape(-1, max_length)
|
| 81 |
+
print(f"Processing {input_ids.size(0)} sequences of length {max_length}")
|
| 82 |
+
|
| 83 |
+
# Run inference
|
| 84 |
+
print("Running inference...")
|
| 85 |
+
_ = model.generate(
|
| 86 |
+
prompt_token_ids=input_ids.tolist(),
|
| 87 |
+
sampling_params=SamplingParams(max_tokens=1)
|
| 88 |
+
)
|
| 89 |
+
|
| 90 |
+
# Save results
|
| 91 |
+
output_path = os.path.join(SAVE_FOLDER, f'activation.{lang}.train.{model_name.split("/")[-1]}')
|
| 92 |
+
torch.save({
|
| 93 |
+
'n': l,
|
| 94 |
+
'over_zero': over_zero.cpu(),
|
| 95 |
+
'num_layers': num_layers,
|
| 96 |
+
'intermediate_size': intermediate_size
|
| 97 |
+
}, output_path)
|
| 98 |
+
|
| 99 |
+
print(f"Saved activation counts to {output_path}")
|
| 100 |
+
print(f"Processed {l} tokens total")
|
| 101 |
+
print("Activation analysis complete!")
|
activation_llama_all.py
ADDED
|
@@ -0,0 +1,146 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
#!/usr/bin/env python3
|
| 2 |
+
"""
|
| 3 |
+
Per-neuron activation tracker for LLaMA-2 and Qwen MLP layers.
|
| 4 |
+
Runs on a fixed set of models and input IDs.
|
| 5 |
+
"""
|
| 6 |
+
|
| 7 |
+
import torch
|
| 8 |
+
import os
|
| 9 |
+
from types import MethodType
|
| 10 |
+
from vllm import LLM, SamplingParams # Keep original import since hook logic depends on vLLM
|
| 11 |
+
|
| 12 |
+
# ---------------------- Config ----------------------
|
| 13 |
+
BASE_PATH = "/home/khanh/sla/sla_cpt"
|
| 14 |
+
|
| 15 |
+
RUN_CONFIGS = [
|
| 16 |
+
{
|
| 17 |
+
'name': 'l2-7b-eu',
|
| 18 |
+
'model': f'{BASE_PATH}/llama2_7b_full_basque_corpus_grad_clip_1/checkpoint-10200',
|
| 19 |
+
'ids_path': './ids/l2-7b/id.eu.train.l2-7b',
|
| 20 |
+
'lang': 'eu',
|
| 21 |
+
'type': 'llama'
|
| 22 |
+
},
|
| 23 |
+
{
|
| 24 |
+
'name': 'l2-13b-ga',
|
| 25 |
+
'model': f'{BASE_PATH}/llama2_13b_full_irish_corpus_grad_clip_1/checkpoint-4280',
|
| 26 |
+
'ids_path': '.ids/l2-13b/id.ga.train.l2-13b',
|
| 27 |
+
'lang': 'en',
|
| 28 |
+
'type': 'llama'
|
| 29 |
+
},
|
| 30 |
+
{
|
| 31 |
+
'name': 'q2.5-zh',
|
| 32 |
+
'model': f'{BASE_PATH}/qwen2.5-0.5_full_chinese_corpus_grad_clip_1/checkpoint-7800',
|
| 33 |
+
'ids_path': './ids/qwen2.5-0.5/id.zh.train.qwen2.5-0.5',
|
| 34 |
+
'lang': 'zh',
|
| 35 |
+
'type': 'qwen'
|
| 36 |
+
},
|
| 37 |
+
{
|
| 38 |
+
'name': 'q2.5-ga',
|
| 39 |
+
'model': f'{BASE_PATH}/qwen2.5-0.5_full_english_corpus_grad_clip_1/checkpoint-3231',
|
| 40 |
+
'ids_path': './ids/qwen2.5-0.5/id.en.train.qwen2.5-0.5',
|
| 41 |
+
'lang': 'ga',
|
| 42 |
+
'type': 'qwen'
|
| 43 |
+
},
|
| 44 |
+
{
|
| 45 |
+
'name': 'q2.5-en+ga',
|
| 46 |
+
'model': f'{BASE_PATH}/qwen2.5-0.5_full_english_corpus_grad_clip_1/checkpoint-3231',
|
| 47 |
+
'ids_path': './ids/qwen2.5-0.5/id.en+ga.train.qwen2.5-0.5',
|
| 48 |
+
'lang': 'ga',
|
| 49 |
+
'type': 'qwen'
|
| 50 |
+
}
|
| 51 |
+
]
|
| 52 |
+
|
| 53 |
+
SAVE_FOLDER = "new_activations"
|
| 54 |
+
os.makedirs(SAVE_FOLDER, exist_ok=True)
|
| 55 |
+
|
| 56 |
+
# ---------------------- Hook Functions ----------------------
|
| 57 |
+
def make_llama_hook(idx):
|
| 58 |
+
def llama_forward(self, x):
|
| 59 |
+
gate_up, _ = self.gate_up_proj(x) # l, 2i
|
| 60 |
+
i = gate_up.size(-1)
|
| 61 |
+
gate_up[:, : i // 2] = torch.nn.SiLU()(gate_up[:, : i // 2])
|
| 62 |
+
activation = gate_up[:, : i // 2].float() # l, i
|
| 63 |
+
over_zero[idx, :] += (activation > 0).sum(dim=(0))
|
| 64 |
+
x = gate_up[:, : i // 2] * gate_up[:, i // 2 :]
|
| 65 |
+
x, _ = self.down_proj(x)
|
| 66 |
+
return x
|
| 67 |
+
return llama_forward
|
| 68 |
+
|
| 69 |
+
def make_qwen_hook(idx):
|
| 70 |
+
def qwen_forward(self, x):
|
| 71 |
+
gate_up, _ = self.gate_up_proj(x) # (s, 2h)
|
| 72 |
+
intermediate_size = gate_up.size(-1) // 2
|
| 73 |
+
gate = gate_up[..., :intermediate_size] # (s, h)
|
| 74 |
+
up = gate_up[..., intermediate_size:] # (s, h)
|
| 75 |
+
gate_activation = torch.nn.functional.silu(gate)
|
| 76 |
+
over_zero[idx, :] += (gate_activation > 0).sum(dim=(0))
|
| 77 |
+
x, _ = self.down_proj(gate_activation * up)
|
| 78 |
+
return x
|
| 79 |
+
return qwen_forward
|
| 80 |
+
|
| 81 |
+
# ---------------------- Run All Configs ----------------------
|
| 82 |
+
for config in RUN_CONFIGS:
|
| 83 |
+
model_name = config['model']
|
| 84 |
+
lang = config['lang']
|
| 85 |
+
ids_path = config['ids_path']
|
| 86 |
+
save_name = config.get('name', model_name)
|
| 87 |
+
model_type = config.get('type', 'llama') # default to 'llama'
|
| 88 |
+
|
| 89 |
+
print(f"\n=== Processing model: {model_name}, lang: {lang}, type: {model_type} ===")
|
| 90 |
+
|
| 91 |
+
# Load model
|
| 92 |
+
model = LLM(
|
| 93 |
+
model=model_name,
|
| 94 |
+
tensor_parallel_size=1,
|
| 95 |
+
enforce_eager=True,
|
| 96 |
+
trust_remote_code=True
|
| 97 |
+
)
|
| 98 |
+
|
| 99 |
+
max_length = model.llm_engine.model_config.max_model_len
|
| 100 |
+
num_layers = model.llm_engine.model_config.hf_config.num_hidden_layers
|
| 101 |
+
intermediate_size = model.llm_engine.model_config.hf_config.intermediate_size
|
| 102 |
+
|
| 103 |
+
print(f"Layers: {num_layers}, Intermediate size: {intermediate_size}, Max length: {max_length}")
|
| 104 |
+
|
| 105 |
+
# Setup activation tracker
|
| 106 |
+
over_zero = torch.zeros(num_layers, intermediate_size, dtype=torch.int32).to('cuda')
|
| 107 |
+
|
| 108 |
+
# Hook MLP layers
|
| 109 |
+
for i in range(num_layers):
|
| 110 |
+
mlp = model.llm_engine.model_executor.driver_worker.model_runner.model.model.layers[i].mlp
|
| 111 |
+
if model_type == 'llama':
|
| 112 |
+
mlp.forward = MethodType(make_llama_hook(i), mlp)
|
| 113 |
+
elif model_type == 'qwen':
|
| 114 |
+
mlp.forward = MethodType(make_qwen_hook(i), mlp)
|
| 115 |
+
else:
|
| 116 |
+
raise ValueError(f"Unknown model type: {model_type}")
|
| 117 |
+
|
| 118 |
+
# Load input IDs
|
| 119 |
+
print("Loading IDs...")
|
| 120 |
+
ids = torch.load(ids_path)
|
| 121 |
+
print(f"ID shape: {ids.shape}")
|
| 122 |
+
|
| 123 |
+
l = ids.size(0)
|
| 124 |
+
l = min(l, 99999744) // max_length * max_length
|
| 125 |
+
input_ids = ids[:l].reshape(-1, max_length)
|
| 126 |
+
print(f"Processing {input_ids.size(0)} sequences of length {max_length}")
|
| 127 |
+
|
| 128 |
+
# Run inference
|
| 129 |
+
print("Running inference...")
|
| 130 |
+
_ = model.generate(
|
| 131 |
+
prompt_token_ids=input_ids.tolist(),
|
| 132 |
+
sampling_params=SamplingParams(max_tokens=1)
|
| 133 |
+
)
|
| 134 |
+
|
| 135 |
+
# Save results
|
| 136 |
+
output_path = os.path.join(SAVE_FOLDER, f'activation.{lang}.train.{save_name}.pt')
|
| 137 |
+
torch.save({
|
| 138 |
+
'n': l,
|
| 139 |
+
'over_zero': over_zero.cpu(),
|
| 140 |
+
'num_layers': num_layers,
|
| 141 |
+
'intermediate_size': intermediate_size
|
| 142 |
+
}, output_path)
|
| 143 |
+
|
| 144 |
+
print(f"Saved activation counts to {output_path}")
|
| 145 |
+
print(f"Processed {l} tokens total")
|
| 146 |
+
print("Activation analysis complete!")
|
activation_qwen.py
ADDED
|
@@ -0,0 +1,109 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
#!/usr/bin/env python3
|
| 2 |
+
"""
|
| 3 |
+
Simple per-neuron activation tracker for Qwen2 MLP layers using vLLM.
|
| 4 |
+
"""
|
| 5 |
+
|
| 6 |
+
import argparse
|
| 7 |
+
from types import MethodType
|
| 8 |
+
import torch
|
| 9 |
+
import os
|
| 10 |
+
from vllm import LLM, SamplingParams
|
| 11 |
+
|
| 12 |
+
# ---------------------- Arguments ----------------------
|
| 13 |
+
parser = argparse.ArgumentParser()
|
| 14 |
+
parser.add_argument("-m", "--model", type=str, required=True, help="HF model ID or local folder path")
|
| 15 |
+
parser.add_argument("-l", "--lang", type=str, required=True, help="Language code for dataset")
|
| 16 |
+
parser.add_argument("--save-folder", type=str, default="data", help="Folder to save activation results")
|
| 17 |
+
args = parser.parse_args()
|
| 18 |
+
|
| 19 |
+
# ---------------------- Ensure Save Folder Exists ----------------------
|
| 20 |
+
os.makedirs(args.save_folder, exist_ok=True)
|
| 21 |
+
|
| 22 |
+
# ---------------------- Setup Model ----------------------
|
| 23 |
+
model = LLM(
|
| 24 |
+
model=args.model,
|
| 25 |
+
tensor_parallel_size=1,
|
| 26 |
+
enforce_eager=True,
|
| 27 |
+
trust_remote_code=True
|
| 28 |
+
)
|
| 29 |
+
|
| 30 |
+
max_length = model.llm_engine.model_config.max_model_len
|
| 31 |
+
num_layers = model.llm_engine.model_config.hf_config.num_hidden_layers
|
| 32 |
+
intermediate_size = model.llm_engine.model_config.hf_config.intermediate_size
|
| 33 |
+
|
| 34 |
+
print(f"Model: {args.model}")
|
| 35 |
+
print(f"Layers: {num_layers}, Intermediate size: {intermediate_size}, Max length: {max_length}")
|
| 36 |
+
|
| 37 |
+
# ---------------------- Setup Tracker ----------------------
|
| 38 |
+
over_zero = torch.zeros(num_layers, intermediate_size, dtype=torch.int32).to('cuda')
|
| 39 |
+
|
| 40 |
+
def make_qwen_hook(idx):
|
| 41 |
+
def qwen_forward(self, x):
|
| 42 |
+
"""
|
| 43 |
+
x: (s, d)
|
| 44 |
+
s: number of tokens
|
| 45 |
+
d: number of features for each tokens
|
| 46 |
+
"""
|
| 47 |
+
|
| 48 |
+
# Qwen2 uses fused gate_up_proj that returns both gate and up in one tensor
|
| 49 |
+
# (s, 2h)
|
| 50 |
+
gate_up, _ = self.gate_up_proj(x)
|
| 51 |
+
|
| 52 |
+
# Split the concatenated gate and up projections
|
| 53 |
+
intermediate_size = gate_up.size(-1) // 2
|
| 54 |
+
# (s, h)
|
| 55 |
+
gate = gate_up[..., :intermediate_size]
|
| 56 |
+
# (s, h)
|
| 57 |
+
up = gate_up[..., intermediate_size:]
|
| 58 |
+
|
| 59 |
+
# Apply activation function to gate
|
| 60 |
+
# (s, h)
|
| 61 |
+
gate_activation = torch.nn.functional.silu(gate)
|
| 62 |
+
|
| 63 |
+
# Track activations > 0
|
| 64 |
+
over_zero[idx, :] += (gate_activation > 0).sum(dim=(0)) # (h)
|
| 65 |
+
|
| 66 |
+
# Complete forward pass: gate * up -> down projection
|
| 67 |
+
x, _ = self.down_proj(gate_activation * up)
|
| 68 |
+
return x
|
| 69 |
+
|
| 70 |
+
return qwen_forward
|
| 71 |
+
|
| 72 |
+
# ---------------------- Hook MLP Layers ----------------------
|
| 73 |
+
print("Setting up activation hooks...")
|
| 74 |
+
for i in range(num_layers):
|
| 75 |
+
mlp = model.llm_engine.model_executor.driver_worker.model_runner.model.model.layers[i].mlp
|
| 76 |
+
mlp.forward = MethodType(make_qwen_hook(i), mlp)
|
| 77 |
+
|
| 78 |
+
# ---------------------- Load Data ----------------------
|
| 79 |
+
print("Loading ids...")
|
| 80 |
+
ids = torch.load(f'../ids/qwen2.5-0.5/id.{args.lang}.train.qwen2.5-0.5')
|
| 81 |
+
print(f"ID shape {ids.shape}")
|
| 82 |
+
|
| 83 |
+
# Process sequences
|
| 84 |
+
l = ids.size(0)
|
| 85 |
+
l = min(l, 99999744) // max_length * max_length
|
| 86 |
+
input_ids = ids[:l].reshape(-1, max_length)
|
| 87 |
+
print(f"Processing {input_ids.size(0)} sequences of length {max_length}")
|
| 88 |
+
|
| 89 |
+
# ---------------------- Run Inference ----------------------
|
| 90 |
+
print("Running inference...")
|
| 91 |
+
output = model.generate(
|
| 92 |
+
prompt_token_ids=input_ids.tolist(),
|
| 93 |
+
sampling_params=SamplingParams(max_tokens=1)
|
| 94 |
+
)
|
| 95 |
+
|
| 96 |
+
# ---------------------- Save Results ----------------------
|
| 97 |
+
result = {
|
| 98 |
+
'n': l,
|
| 99 |
+
'over_zero': over_zero.to('cpu'),
|
| 100 |
+
'num_layers': num_layers,
|
| 101 |
+
'intermediate_size': intermediate_size
|
| 102 |
+
}
|
| 103 |
+
|
| 104 |
+
output_path = os.path.join(args.save_folder, f'activation.{args.lang}.train.qwen-{args.model.split("/")[-1]}')
|
| 105 |
+
torch.save(result, output_path)
|
| 106 |
+
|
| 107 |
+
print(f"Saved activation counts to {output_path}")
|
| 108 |
+
print(f"Processed {l} tokens total")
|
| 109 |
+
print("Activation analysis complete!")
|
activation_qwen_all.py
ADDED
|
@@ -0,0 +1,126 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
#!/usr/bin/env python3
|
| 2 |
+
"""
|
| 3 |
+
Simple per-neuron activation tracker for LLaMA-2 MLP layers.
|
| 4 |
+
Runs on a fixed set of models and input IDs.
|
| 5 |
+
"""
|
| 6 |
+
|
| 7 |
+
import torch
|
| 8 |
+
import os
|
| 9 |
+
from types import MethodType
|
| 10 |
+
from vllm import LLM, SamplingParams # Keep your original import since hook logic depends on vLLM
|
| 11 |
+
|
| 12 |
+
# ---------------------- Config ----------------------
|
| 13 |
+
|
| 14 |
+
BASE_PATH = "/home/khanh/sla/sla_cpt"
|
| 15 |
+
|
| 16 |
+
RUN_CONFIGS = [
|
| 17 |
+
{
|
| 18 |
+
'name': 'q2.5-zh', # custom name for output file
|
| 19 |
+
'model': f'{BASE_PATH}/llama2_7b_full_basque_corpus_grad_clip_1/checkpoint-10200', # replace with LLaMA2 HF model if needed
|
| 20 |
+
'ids_path': './ids/l2-7b/id.eu.train.l2-7b',
|
| 21 |
+
'lang': 'eu'
|
| 22 |
+
},
|
| 23 |
+
{
|
| 24 |
+
'name': 'q2.5-ga',
|
| 25 |
+
'model': f'{BASE_PATH}/llama2_13b_full_irish_corpus_grad_clip_1/checkpoint-4280',
|
| 26 |
+
'ids_path': '.ids/l2-13b/id.ga.train.l2-13b',
|
| 27 |
+
'lang': 'en'
|
| 28 |
+
}
|
| 29 |
+
]
|
| 30 |
+
|
| 31 |
+
SAVE_FOLDER = "new_activations"
|
| 32 |
+
os.makedirs(SAVE_FOLDER, exist_ok=True)
|
| 33 |
+
|
| 34 |
+
# ---------------------- Original Hook Function ----------------------
|
| 35 |
+
def make_qwen_hook(idx):
|
| 36 |
+
def qwen_forward(self, x):
|
| 37 |
+
"""
|
| 38 |
+
x: (s, d)
|
| 39 |
+
s: number of tokens
|
| 40 |
+
d: number of features for each tokens
|
| 41 |
+
"""
|
| 42 |
+
|
| 43 |
+
# Qwen2 uses fused gate_up_proj that returns both gate and up in one tensor
|
| 44 |
+
# (s, 2h)
|
| 45 |
+
gate_up, _ = self.gate_up_proj(x)
|
| 46 |
+
|
| 47 |
+
# Split the concatenated gate and up projections
|
| 48 |
+
intermediate_size = gate_up.size(-1) // 2
|
| 49 |
+
# (s, h)
|
| 50 |
+
gate = gate_up[..., :intermediate_size]
|
| 51 |
+
# (s, h)
|
| 52 |
+
up = gate_up[..., intermediate_size:]
|
| 53 |
+
|
| 54 |
+
# Apply activation function to gate
|
| 55 |
+
# (s, h)
|
| 56 |
+
gate_activation = torch.nn.functional.silu(gate)
|
| 57 |
+
|
| 58 |
+
# Track activations > 0
|
| 59 |
+
over_zero[idx, :] += (gate_activation > 0).sum(dim=(0)) # (h)
|
| 60 |
+
|
| 61 |
+
# Complete forward pass: gate * up -> down projection
|
| 62 |
+
x, _ = self.down_proj(gate_activation * up)
|
| 63 |
+
return x
|
| 64 |
+
|
| 65 |
+
return qwen_forward
|
| 66 |
+
|
| 67 |
+
# ---------------------- Run All Configs ----------------------
|
| 68 |
+
for config in RUN_CONFIGS:
|
| 69 |
+
model_name = config['model']
|
| 70 |
+
lang = config['lang']
|
| 71 |
+
ids_path = config['ids_path']
|
| 72 |
+
save_name = config.get('name', model_name) # use 'name' key if present, otherwise fallback to model_name
|
| 73 |
+
|
| 74 |
+
print(f"\n=== Processing model: {model_name}, lang: {lang} ===")
|
| 75 |
+
|
| 76 |
+
# Load model
|
| 77 |
+
model = LLM(
|
| 78 |
+
model=model_name,
|
| 79 |
+
tensor_parallel_size=1,
|
| 80 |
+
enforce_eager=True,
|
| 81 |
+
trust_remote_code=True
|
| 82 |
+
)
|
| 83 |
+
|
| 84 |
+
max_length = model.llm_engine.model_config.max_model_len
|
| 85 |
+
num_layers = model.llm_engine.model_config.hf_config.num_hidden_layers
|
| 86 |
+
intermediate_size = model.llm_engine.model_config.hf_config.intermediate_size
|
| 87 |
+
|
| 88 |
+
print(f"Layers: {num_layers}, Intermediate size: {intermediate_size}, Max length: {max_length}")
|
| 89 |
+
|
| 90 |
+
# Setup activation tracker
|
| 91 |
+
over_zero = torch.zeros(num_layers, intermediate_size, dtype=torch.int32).to('cuda')
|
| 92 |
+
|
| 93 |
+
# Hook MLP layers
|
| 94 |
+
for i in range(num_layers):
|
| 95 |
+
mlp = model.llm_engine.model_executor.driver_worker.model_runner.model.model.layers[i].mlp
|
| 96 |
+
mlp.forward = MethodType(make_qwen_hook(i), mlp)
|
| 97 |
+
|
| 98 |
+
# Load input IDs
|
| 99 |
+
print("Loading IDs...")
|
| 100 |
+
ids = torch.load(ids_path)
|
| 101 |
+
print(f"ID shape: {ids.shape}")
|
| 102 |
+
|
| 103 |
+
l = ids.size(0)
|
| 104 |
+
l = min(l, 99999744) // max_length * max_length
|
| 105 |
+
input_ids = ids[:l].reshape(-1, max_length)
|
| 106 |
+
print(f"Processing {input_ids.size(0)} sequences of length {max_length}")
|
| 107 |
+
|
| 108 |
+
# Run inference
|
| 109 |
+
print("Running inference...")
|
| 110 |
+
_ = model.generate(
|
| 111 |
+
prompt_token_ids=input_ids.tolist(),
|
| 112 |
+
sampling_params=SamplingParams(max_tokens=1)
|
| 113 |
+
)
|
| 114 |
+
|
| 115 |
+
# Save results using the 'name' key
|
| 116 |
+
output_path = os.path.join(SAVE_FOLDER, f'activation.{lang}.train.{save_name}.pt')
|
| 117 |
+
torch.save({
|
| 118 |
+
'n': l,
|
| 119 |
+
'over_zero': over_zero.cpu(),
|
| 120 |
+
'num_layers': num_layers,
|
| 121 |
+
'intermediate_size': intermediate_size
|
| 122 |
+
}, output_path)
|
| 123 |
+
|
| 124 |
+
print(f"Saved activation counts to {output_path}")
|
| 125 |
+
print(f"Processed {l} tokens total")
|
| 126 |
+
print("Activation analysis complete!")
|
activation_single.py
ADDED
|
@@ -0,0 +1,76 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
import argparse
|
| 2 |
+
import os
|
| 3 |
+
from types import MethodType
|
| 4 |
+
|
| 5 |
+
import torch
|
| 6 |
+
from vllm import LLM, SamplingParams
|
| 7 |
+
|
| 8 |
+
|
| 9 |
+
parser = argparse.ArgumentParser()
|
| 10 |
+
parser.add_argument("-m", "--model", type=str, default="meta-llama/Llama-2-7b-hf")
|
| 11 |
+
parser.add_argument("-i", "--id_path", type=str, required=True)
|
| 12 |
+
parser.add_argument("-t", "--type", type=str, default="llama")
|
| 13 |
+
parser.add_argument("-s", "--save_folder", type=str, required=True)
|
| 14 |
+
parser.add_argument("-n", "--name", type=str, required=True)
|
| 15 |
+
args = parser.parse_args()
|
| 16 |
+
|
| 17 |
+
model = LLM(model=args.model, tensor_parallel_size=torch.cuda.device_count(), enforce_eager=True)
|
| 18 |
+
|
| 19 |
+
max_length = model.llm_engine.model_config.max_model_len
|
| 20 |
+
num_layers = model.llm_engine.model_config.hf_config.num_hidden_layers
|
| 21 |
+
intermediate_size = model.llm_engine.model_config.hf_config.intermediate_size
|
| 22 |
+
|
| 23 |
+
over_zero = torch.zeros(num_layers, intermediate_size, dtype=torch.int32).to('cuda')
|
| 24 |
+
|
| 25 |
+
def extract_lang(id_path):
|
| 26 |
+
parts = id_path.split('/')
|
| 27 |
+
id = parts[-1]
|
| 28 |
+
|
| 29 |
+
lang = id.split('.')[1]
|
| 30 |
+
return lang
|
| 31 |
+
|
| 32 |
+
def factory(idx):
|
| 33 |
+
def llama_forward(self, x):
|
| 34 |
+
gate_up, _ = self.gate_up_proj(x) # l, 2i
|
| 35 |
+
i = gate_up.size(-1)
|
| 36 |
+
gate_up[:, : i // 2] = torch.nn.SiLU()(gate_up[:, : i // 2])
|
| 37 |
+
activation = gate_up[:, : i // 2].float() # l, i
|
| 38 |
+
over_zero[idx, :] += (activation > 0).sum(dim=0)
|
| 39 |
+
x = gate_up[:, : i // 2] * gate_up[:, i // 2 :]
|
| 40 |
+
x, _ = self.down_proj(x)
|
| 41 |
+
return x
|
| 42 |
+
|
| 43 |
+
def qwen_forward(self, x):
|
| 44 |
+
gate_up, _ = self.gate_up_proj(x) # (s, 2h)
|
| 45 |
+
intermediate_size = gate_up.size(-1) // 2
|
| 46 |
+
gate = gate_up[..., :intermediate_size] # (s, h)
|
| 47 |
+
up = gate_up[..., intermediate_size:] # (s, h)
|
| 48 |
+
gate_activation = torch.nn.functional.silu(gate)
|
| 49 |
+
over_zero[idx, :] += (gate_activation > 0).sum(dim=0)
|
| 50 |
+
x, _ = self.down_proj(gate_activation * up)
|
| 51 |
+
return x
|
| 52 |
+
|
| 53 |
+
if args.type == 'llama':
|
| 54 |
+
return llama_forward
|
| 55 |
+
else:
|
| 56 |
+
return qwen_forward
|
| 57 |
+
|
| 58 |
+
for i in range(num_layers):
|
| 59 |
+
obj = model.llm_engine.driver_worker.model_runner.model.model.layers[i].mlp
|
| 60 |
+
obj.forward = MethodType(factory(i), obj)
|
| 61 |
+
|
| 62 |
+
lang = args.lang
|
| 63 |
+
|
| 64 |
+
ids = torch.load(args.id_path)
|
| 65 |
+
|
| 66 |
+
l = ids.size(0)
|
| 67 |
+
l = min(l, 99999744) // max_length * max_length
|
| 68 |
+
input_ids = ids[:l].reshape(-1, max_length)
|
| 69 |
+
|
| 70 |
+
output = model.generate(prompt_token_ids=input_ids.tolist(), sampling_params=SamplingParams(max_tokens=1))
|
| 71 |
+
|
| 72 |
+
output = dict(n=l, over_zero=over_zero.to('cpu'))
|
| 73 |
+
|
| 74 |
+
save_path = os.path.join(args.save_folder, f"activation.{lang}.{args.name}.pt")
|
| 75 |
+
|
| 76 |
+
torch.save(output, save_path)
|
activation_single.sh
ADDED
|
@@ -0,0 +1,55 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
#!/bin/bash
|
| 2 |
+
#SBATCH --job-name=activation_single # Job name
|
| 3 |
+
#SBATCH --output=activation_single.log # Standard output log
|
| 4 |
+
#SBATCH --error=activation_single.log # Standard error log
|
| 5 |
+
#SBATCH --ntasks=1 # Number of tasks (processes)
|
| 6 |
+
#SBATCH --cpus-per-task=32 # Number of CPU cores
|
| 7 |
+
#SBATCH --mem=256G # Memory
|
| 8 |
+
#SBATCH --gres=gpu:1 # Number of GPUs
|
| 9 |
+
#SBATCH --partition=physical-gpu # Partition/queue name
|
| 10 |
+
|
| 11 |
+
#!/usr/bin/env bash
|
| 12 |
+
|
| 13 |
+
BASE_PATH="/home/khanh/sla/sla_cpt"
|
| 14 |
+
ID_BASE_PATH="./oscar_ids"
|
| 15 |
+
ACTIVATION_BASE_PATH="./new_new_activations"
|
| 16 |
+
|
| 17 |
+
# --- Run Configs (clean style) ---
|
| 18 |
+
# run_l2_13b=(
|
| 19 |
+
# name="l2-13b"
|
| 20 |
+
# model="$BASE_PATH/uccix/checkpoint-4280"
|
| 21 |
+
# type="llama"
|
| 22 |
+
# ids=(
|
| 23 |
+
# "$ID_BASE_PATH/l2-13b/id.ga.train.l2-13b:ga"
|
| 24 |
+
# "$ID_BASE_PATH/l2-13b/id.en.train.l2-13b:en"
|
| 25 |
+
# )
|
| 26 |
+
# )
|
| 27 |
+
|
| 28 |
+
run_q25_zh=(
|
| 29 |
+
name="q2.5-zh"
|
| 30 |
+
model="$BASE_PATH/qwen2.5-0.5b_english_wiki_750M_chinese_wikipedia_corpus_2e_240925/checkpoint-2944"
|
| 31 |
+
type="qwen"
|
| 32 |
+
ids=(
|
| 33 |
+
"$ID_BASE_PATH/q2.5/id.zh.train.q2.5"
|
| 34 |
+
"$ID_BASE_PATH/q2.5/id.en.train.q2.5"
|
| 35 |
+
)
|
| 36 |
+
)
|
| 37 |
+
|
| 38 |
+
# --- Put them in a list ---
|
| 39 |
+
ALL_RUNS=(run_q25_zh)
|
| 40 |
+
|
| 41 |
+
# --- Loop ---
|
| 42 |
+
for run in "${ALL_RUNS[@]}"; do
|
| 43 |
+
declare -n cfg="$run"
|
| 44 |
+
|
| 45 |
+
echo "=== Running ${cfg[name]} ==="
|
| 46 |
+
for id_path in "${cfg[ids][@]}"; do
|
| 47 |
+
echo " -> $id_path"
|
| 48 |
+
python3 activation_single.py \
|
| 49 |
+
-m "${cfg[model]}" \
|
| 50 |
+
-t "${cfg[type]}" \
|
| 51 |
+
-i "$id_path" \
|
| 52 |
+
-n "${cfg[name]}" \
|
| 53 |
+
-s "$ACTIVATION_BASE_PATH"
|
| 54 |
+
done
|
| 55 |
+
done
|
activations-old/README.md
ADDED
|
@@ -0,0 +1 @@
|
|
|
|
|
|
|
| 1 |
+
Activations on WIKIPEDIA DATASET
|
activations-old/qwen2.5-0.5b_english_wiki_1.5B_basque_corpus/activation.en.train.qwen-checkpoint-1200
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:9cea277fb7636f145b1bf551617d1ed05fa07e67dac1d95842061d53f4d3c4e9
|
| 3 |
+
size 468248
|
activations-old/qwen2.5-0.5b_english_wiki_1.5B_basque_corpus/activation.en.train.qwen-checkpoint-1500
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:1749441df8a2860e14a41583860cf7fb836534f2b2e181c13e81cb123726667c
|
| 3 |
+
size 468248
|
activations-old/qwen2.5-0.5b_english_wiki_1.5B_basque_corpus/activation.en.train.qwen-checkpoint-1800
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:726d65b3d0020771d34abf07322cc2b9166ba78b13169f403643ca47a6d63880
|
| 3 |
+
size 468248
|
activations-old/qwen2.5-0.5b_english_wiki_1.5B_basque_corpus/activation.en.train.qwen-checkpoint-2100
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:a0bbb84b21288c13c83debf4e979bfb533a2f3b429427da7db76fcec78f12a88
|
| 3 |
+
size 468248
|
activations-old/qwen2.5-0.5b_english_wiki_1.5B_basque_corpus/activation.en.train.qwen-checkpoint-2400
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:fe5cb042ba2bf8008f279aa0647f3fdc60671374b76677f1997308831ca60b73
|
| 3 |
+
size 468248
|
activations-old/qwen2.5-0.5b_english_wiki_1.5B_basque_corpus/activation.en.train.qwen-checkpoint-2700
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:817783e55b07a2703698edf7f1381a92c45ae29d464a3992f4979b821aca0159
|
| 3 |
+
size 468248
|
activations-old/qwen2.5-0.5b_english_wiki_1.5B_basque_corpus/activation.en.train.qwen-checkpoint-300
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:d133b8a1d9c909a8079b91236d7e0ce079cc0fd5f304d9404f17070518700b57
|
| 3 |
+
size 468248
|
activations-old/qwen2.5-0.5b_english_wiki_1.5B_basque_corpus/activation.en.train.qwen-checkpoint-3000
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:1fb497bf1342c7566510502713b8035ef26fb5c5fe19cbc9b72bbbfe7a07dfd0
|
| 3 |
+
size 468248
|
activations-old/qwen2.5-0.5b_english_wiki_1.5B_basque_corpus/activation.en.train.qwen-checkpoint-3231
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:c856e80ecb87701a23f4984429da9b9c0decb14c1ad7d503bbbb524b29fb299f
|
| 3 |
+
size 468248
|
activations-old/qwen2.5-0.5b_english_wiki_1.5B_basque_corpus/activation.en.train.qwen-checkpoint-600
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:8740d524564999015bd919f5a166deab4b03e1fc4c2730564695f4f76abf9ee1
|
| 3 |
+
size 468248
|
activations-old/qwen2.5-0.5b_english_wiki_1.5B_basque_corpus/activation.en.train.qwen-checkpoint-900
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:c5beabf8349f63597e36299db52121e47af81b4e4a02f4439b4c9f34f01817a2
|
| 3 |
+
size 468248
|
activations-old/qwen2.5-0.5b_english_wiki_1.5B_basque_corpus/activation.eu.train.qwen-checkpoint-1200
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:1504808cbf37639effd1664a0bff790d5be28ec3e23b17ed65c12c7d3c08d210
|
| 3 |
+
size 468248
|
activations-old/qwen2.5-0.5b_english_wiki_1.5B_basque_corpus/activation.eu.train.qwen-checkpoint-1500
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:30cce6951837c6d19d484e16603abd0779c2cbcf3f818e0e08827089bb9ffe60
|
| 3 |
+
size 468248
|
activations-old/qwen2.5-0.5b_english_wiki_1.5B_basque_corpus/activation.eu.train.qwen-checkpoint-1800
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:41c625eab5041de7d461b822498f82ce20fbbcceaaa703cb8af5b970f029703a
|
| 3 |
+
size 468248
|
activations-old/qwen2.5-0.5b_english_wiki_1.5B_basque_corpus/activation.eu.train.qwen-checkpoint-2100
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:ad1dad46b038443c45a9df3560ba00e31e68ea85248cf2f2aa2bea0e53390309
|
| 3 |
+
size 468248
|
activations-old/qwen2.5-0.5b_english_wiki_1.5B_basque_corpus/activation.eu.train.qwen-checkpoint-2400
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:c4404b495aa455a0689c951c9da191b87be2ca6c91e4c1614d7d1eeb5da7572f
|
| 3 |
+
size 468248
|
activations-old/qwen2.5-0.5b_english_wiki_1.5B_basque_corpus/activation.eu.train.qwen-checkpoint-2700
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:7104954d86a009d4968d3abb8c302a38789b3abbf1fee277c57541420e547c9e
|
| 3 |
+
size 468248
|
activations-old/qwen2.5-0.5b_english_wiki_1.5B_basque_corpus/activation.eu.train.qwen-checkpoint-300
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:61b0911c5246ed8d8cd93e3bd55771b678e56e91266008205473b01d6d0c58aa
|
| 3 |
+
size 468248
|
activations-old/qwen2.5-0.5b_english_wiki_1.5B_basque_corpus/activation.eu.train.qwen-checkpoint-3000
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:21c10d4298ef08be77d858ff161dec601d9b26b610057a81ae4d3eb0dda9f3f8
|
| 3 |
+
size 468248
|
activations-old/qwen2.5-0.5b_english_wiki_1.5B_basque_corpus/activation.eu.train.qwen-checkpoint-3231
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:a84e27588461f9afcbe2cd08da3b31e983315b25afe43d93806b45f95cdda4c1
|
| 3 |
+
size 468248
|
activations-old/qwen2.5-0.5b_english_wiki_1.5B_basque_corpus/activation.eu.train.qwen-checkpoint-600
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:24d61fe7398bc86d25790a4b4fff6a9bf33f1b2657c3ace3b67f520f6a8e8bbd
|
| 3 |
+
size 468248
|
activations-old/qwen2.5-0.5b_english_wiki_1.5B_basque_corpus/activation.eu.train.qwen-checkpoint-900
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:6fe9a1c39f9ed33536dab2417a976b18abb4aab42b233a99ca558ed61f12e32a
|
| 3 |
+
size 468248
|
activations-old/qwen2.5-0.5b_english_wiki_300M_en_1.5Bbasque_corpus/activation.en.train.qwen-checkpoint-1200
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:021759d151d96f4f10861bae0cea0104e5600361c65ae9e1f135ec88a91c219b
|
| 3 |
+
size 468248
|
activations-old/qwen2.5-0.5b_english_wiki_300M_en_1.5Bbasque_corpus/activation.en.train.qwen-checkpoint-1500
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:9fc65a08a2936427c6abd4342967b9c67d5921641795a5f9de7fc964d9fbd400
|
| 3 |
+
size 468248
|
activations-old/qwen2.5-0.5b_english_wiki_300M_en_1.5Bbasque_corpus/activation.en.train.qwen-checkpoint-1800
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:2fb4f35aeffedb6d1efb928cf587ce44a445f476ca3821d5cb44ad3c7b8dff0e
|
| 3 |
+
size 468248
|
activations-old/qwen2.5-0.5b_english_wiki_300M_en_1.5Bbasque_corpus/activation.en.train.qwen-checkpoint-2100
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:ff03e3081c471e6b71899e1327a9c4160defa427a48cad6c19176fd427b5d504
|
| 3 |
+
size 468248
|
activations-old/qwen2.5-0.5b_english_wiki_300M_en_1.5Bbasque_corpus/activation.en.train.qwen-checkpoint-2400
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:7ad1f7ab916924bb34d53647cf2bd2760b944e8870b49f63ad8056cbd6b69df2
|
| 3 |
+
size 468248
|
activations-old/qwen2.5-0.5b_english_wiki_300M_en_1.5Bbasque_corpus/activation.en.train.qwen-checkpoint-2700
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:6467749237cf7074443d52a4ab7a253426d68dc4caf4dc13c01d917c5e3ddbef
|
| 3 |
+
size 468248
|
activations-old/qwen2.5-0.5b_english_wiki_300M_en_1.5Bbasque_corpus/activation.en.train.qwen-checkpoint-2972
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:95de6afadc9155ee59b66574af57fefbee9950ca9fde08ffc6914cb13f7e9290
|
| 3 |
+
size 468248
|
activations-old/qwen2.5-0.5b_english_wiki_300M_en_1.5Bbasque_corpus/activation.en.train.qwen-checkpoint-300
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:d796bc5f4df79f526568aa817162eae20d29bf6d46ffa3b803893fd305a4b482
|
| 3 |
+
size 468248
|
activations-old/qwen2.5-0.5b_english_wiki_300M_en_1.5Bbasque_corpus/activation.en.train.qwen-checkpoint-600
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:d9cc9a96849fff460e5c5294386a9d7637108234938f705f2313a8e09515d3d2
|
| 3 |
+
size 468248
|
activations-old/qwen2.5-0.5b_english_wiki_300M_en_1.5Bbasque_corpus/activation.en.train.qwen-checkpoint-900
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:b7e1ea9a62d6cb3f6f69d245b02e89c1f0f9e01d8c940e4d3fc1186fa7db8e86
|
| 3 |
+
size 468248
|
activations-old/qwen2.5-0.5b_english_wiki_300M_en_1.5Bbasque_corpus/activation.eu.train.qwen-checkpoint-1200
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:300abc0c19cee744c435c9f596152cba7d6d591be831b1a633906371d19eabc8
|
| 3 |
+
size 468248
|
activations-old/qwen2.5-0.5b_english_wiki_300M_en_1.5Bbasque_corpus/activation.eu.train.qwen-checkpoint-1500
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:090ce2e0d3ab3aa6a197af9e47bfcc44f2490702a0cd5ccca93c54a5cbb24f70
|
| 3 |
+
size 468248
|
activations-old/qwen2.5-0.5b_english_wiki_300M_en_1.5Bbasque_corpus/activation.eu.train.qwen-checkpoint-1800
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:d1bfb66fd295ceac5b29d31018905e6a5ecf9530bcd4f3bab8bfd77a8d789b5d
|
| 3 |
+
size 468248
|
activations-old/qwen2.5-0.5b_english_wiki_300M_en_1.5Bbasque_corpus/activation.eu.train.qwen-checkpoint-2100
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:267ff05b6ccd58d67de3874ca3a56ab8adea14389d5f1cd931a8367a735f9012
|
| 3 |
+
size 468248
|
activations-old/qwen2.5-0.5b_english_wiki_300M_en_1.5Bbasque_corpus/activation.eu.train.qwen-checkpoint-2400
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:f791657a1175690e59daa395eab1e2d73a98dec3c03585d8aeaef250b5bf5ba5
|
| 3 |
+
size 468248
|