Update README.md
Browse files
README.md
CHANGED
|
@@ -5,9 +5,9 @@ widget:
|
|
| 5 |
- text: "作为电子extra0的平台，京东绝对是领先者。如今的刘强extra1已经是身价过extra2的老板。"
|
| 6 |
|
| 7 |
|
| 8 |
-
|
| 9 |
---
|
| 10 |
|
|
|
|
| 11 |
# Chinese T5 Version 1.1
|
| 12 |
|
| 13 |
## Model description
|
|
@@ -61,7 +61,7 @@ python3 preprocess.py --corpus_path corpora/cluecorpussmall.txt \
|
|
| 61 |
--vocab_path models/google_zh_with_sentinel_vocab.txt \
|
| 62 |
--dataset_path cluecorpussmall_t5-v1_1_seq128_dataset.pt \
|
| 63 |
--processes_num 32 --seq_length 128 \
|
| 64 |
-
--dynamic_masking --
|
| 65 |
```
|
| 66 |
|
| 67 |
```
|
|
@@ -72,10 +72,7 @@ python3 pretrain.py --dataset_path cluecorpussmall_t5-v1_1_seq128_dataset.pt \
|
|
| 72 |
--world_size 8 --gpu_ranks 0 1 2 3 4 5 6 7 \
|
| 73 |
--total_steps 1000000 --save_checkpoint_steps 100000 --report_steps 50000 \
|
| 74 |
--learning_rate 1e-3 --batch_size 64 \
|
| 75 |
-
--span_masking --span_geo_prob 0.3 --span_max_length 5
|
| 76 |
-
--embedding word --relative_position_embedding --remove_embedding_layernorm --tgt_embedding word \
|
| 77 |
-
--encoder transformer --mask fully_visible --layernorm_positioning pre \
|
| 78 |
-
--feed_forward gated --decoder transformer --target t5
|
| 79 |
```
|
| 80 |
|
| 81 |
Stage2:
|
|
@@ -85,7 +82,7 @@ python3 preprocess.py --corpus_path corpora/cluecorpussmall.txt \
|
|
| 85 |
--vocab_path models/google_zh_with_sentinel_vocab.txt \
|
| 86 |
--dataset_path cluecorpussmall_t5-v1_1_seq512_dataset.pt \
|
| 87 |
--processes_num 32 --seq_length 512 \
|
| 88 |
-
--dynamic_masking --
|
| 89 |
```
|
| 90 |
|
| 91 |
```
|
|
@@ -97,10 +94,7 @@ python3 pretrain.py --dataset_path cluecorpussmall_t5-v1_1_seq512_dataset.pt \
|
|
| 97 |
--world_size 8 --gpu_ranks 0 1 2 3 4 5 6 7 \
|
| 98 |
--total_steps 250000 --save_checkpoint_steps 50000 --report_steps 10000 \
|
| 99 |
--learning_rate 5e-4 --batch_size 16 \
|
| 100 |
-
--span_masking --span_geo_prob 0.3 --span_max_length 5
|
| 101 |
-
--embedding word --relative_position_embedding --remove_embedding_layernorm --tgt_embedding word \
|
| 102 |
-
--encoder transformer --mask fully_visible --layernorm_positioning pre \
|
| 103 |
-
--feed_forward gated --decoder transformer --target t5
|
| 104 |
```
|
| 105 |
|
| 106 |
Finally, we convert the pre-trained model into Huggingface's format:
|
|
|
|
| 5 |
- text: "作为电子extra0的平台，京东绝对是领先者。如今的刘强extra1已经是身价过extra2的老板。"
|
| 6 |
|
| 7 |
|
|
|
|
| 8 |
---
|
| 9 |
|
| 10 |
+
|
| 11 |
# Chinese T5 Version 1.1
|
| 12 |
|
| 13 |
## Model description
|
|
|
|
| 61 |
--vocab_path models/google_zh_with_sentinel_vocab.txt \
|
| 62 |
--dataset_path cluecorpussmall_t5-v1_1_seq128_dataset.pt \
|
| 63 |
--processes_num 32 --seq_length 128 \
|
| 64 |
+
--dynamic_masking --data_processor t5
|
| 65 |
```
|
| 66 |
|
| 67 |
```
|
|
|
|
| 72 |
--world_size 8 --gpu_ranks 0 1 2 3 4 5 6 7 \
|
| 73 |
--total_steps 1000000 --save_checkpoint_steps 100000 --report_steps 50000 \
|
| 74 |
--learning_rate 1e-3 --batch_size 64 \
|
| 75 |
+
--span_masking --span_geo_prob 0.3 --span_max_length 5
|
|
|
|
|
|
|
|
|
|
| 76 |
```
|
| 77 |
|
| 78 |
Stage2:
|
|
|
|
| 82 |
--vocab_path models/google_zh_with_sentinel_vocab.txt \
|
| 83 |
--dataset_path cluecorpussmall_t5-v1_1_seq512_dataset.pt \
|
| 84 |
--processes_num 32 --seq_length 512 \
|
| 85 |
+
--dynamic_masking --data_processor t5
|
| 86 |
```
|
| 87 |
|
| 88 |
```
|
|
|
|
| 94 |
--world_size 8 --gpu_ranks 0 1 2 3 4 5 6 7 \
|
| 95 |
--total_steps 250000 --save_checkpoint_steps 50000 --report_steps 10000 \
|
| 96 |
--learning_rate 5e-4 --batch_size 16 \
|
| 97 |
+
--span_masking --span_geo_prob 0.3 --span_max_length 5
|
|
|
|
|
|
|
|
|
|
| 98 |
```
|
| 99 |
|
| 100 |
Finally, we convert the pre-trained model into Huggingface's format:
|