Update README.md
Browse files
README.md
CHANGED
@@ -20,11 +20,12 @@ tags:
|
|
20 |
|
21 |
### 数据集
|
22 |
|
23 |
-
以 m-a-p/neo_sft_phase2
|
24 |
|
25 |
1. REILX/neo_sft_phase2_conversations
|
26 |
2. REILX/neo_sft_phase2_multi
|
27 |
3. REILX/neo_sft_phase2_single
|
|
|
28 |
|
29 |
### 数据集构建规则
|
30 |
|
@@ -55,6 +56,16 @@ tags:
|
|
55 |
4. 将该“conversation”的“gpt”的“value”作为“output”。
|
56 |
5. “input”可为空白,亦可注入适当的提示信息。
|
57 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
58 |
### 训练参数
|
59 |
REILX/neo_sft_phase2_conversations</br>
|
60 |
The following hyperparameters were used during training:
|
@@ -105,6 +116,22 @@ REILX/neo_sft_phase2_single</br>
|
|
105 |
- lr_scheduler_warmup_ratio: 0.1
|
106 |
- num_epochs: 5.0
|
107 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
108 |
### 损失图
|
109 |
REILX/neo_sft_phase2_conversations</br>
|
110 |
<img src="./neo_sft_phase2_conversations/training_loss.png" alt="neo_sft_phase2_conversations_loss" width="60%">
|
@@ -113,4 +140,8 @@ REILX/neo_sft_phase2_multi</br>
|
|
113 |
<img src="./neo_sft_phase2_multi/training_loss.png" alt="neo_sft_phase2_multi_loss" width="60%">
|
114 |
|
115 |
REILX/neo_sft_phase2_single</br>
|
116 |
-
<img src="./neo_sft_phase2_single/training_loss.png" alt="neo_sft_phase2_single_loss" width="60%">
|
|
|
|
|
|
|
|
|
|
20 |
|
21 |
### 数据集
|
22 |
|
23 |
+
以 m-a-p/neo_sft_phase2 数据集为基石,构建了四个子数据集,分别如下:
|
24 |
|
25 |
1. REILX/neo_sft_phase2_conversations
|
26 |
2. REILX/neo_sft_phase2_multi
|
27 |
3. REILX/neo_sft_phase2_single
|
28 |
+
4. REILX/neo_sft_phase2_all_pair
|
29 |
|
30 |
### 数据集构建规则
|
31 |
|
|
|
56 |
4. 将该“conversation”的“gpt”的“value”作为“output”。
|
57 |
5. “input”可为空白,亦可注入适当的提示信息。
|
58 |
|
59 |
+
**REILX/neo_sft_phase2_all_pair**
|
60 |
+
|
61 |
+
* **具体步骤:**
|
62 |
+
1. 输入为一个json文件,遍历每一个conversations
|
63 |
+
2. conversations包含多轮对话,需要按照对应的轮数构成新数据集
|
64 |
+
3. 比如1、2轮构成一个jsonl的一行,3、4构成一行,5、6构成一行等等等,直到完整的使用结束conversations
|
65 |
+
4. 将该“conversation”的“human”的“value”作为“instruction”
|
66 |
+
5. 将该“conversation”的“gpt”的“value”作为“output”
|
67 |
+
4. “input”可为空白,亦可注入适当的提示信息。
|
68 |
+
|
69 |
### 训练参数
|
70 |
REILX/neo_sft_phase2_conversations</br>
|
71 |
The following hyperparameters were used during training:
|
|
|
116 |
- lr_scheduler_warmup_ratio: 0.1
|
117 |
- num_epochs: 5.0
|
118 |
|
119 |
+
REILX/neo_sft_phase2_all_pair</br>
|
120 |
+
- learning_rate: 2e-05
|
121 |
+
- train_batch_size: 1
|
122 |
+
- eval_batch_size: 8
|
123 |
+
- cutoff_len:4096
|
124 |
+
- seed: 42
|
125 |
+
- distributed_type: multi-GPU
|
126 |
+
- num_devices: 8
|
127 |
+
- gradient_accumulation_steps: 8
|
128 |
+
- total_train_batch_size: 64
|
129 |
+
- total_eval_batch_size: 64
|
130 |
+
- optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
|
131 |
+
- lr_scheduler_type: cosine
|
132 |
+
- lr_scheduler_warmup_ratio: 0.1
|
133 |
+
- num_epochs: 5.0
|
134 |
+
|
135 |
### 损失图
|
136 |
REILX/neo_sft_phase2_conversations</br>
|
137 |
<img src="./neo_sft_phase2_conversations/training_loss.png" alt="neo_sft_phase2_conversations_loss" width="60%">
|
|
|
140 |
<img src="./neo_sft_phase2_multi/training_loss.png" alt="neo_sft_phase2_multi_loss" width="60%">
|
141 |
|
142 |
REILX/neo_sft_phase2_single</br>
|
143 |
+
<img src="./neo_sft_phase2_single/training_loss.png" alt="neo_sft_phase2_single_loss" width="60%">
|
144 |
+
|
145 |
+
REILX/neo_sft_phase2_all_pair</br>
|
146 |
+
<!-- ![neo_sft_phase2_single_loss](./neo_sft_phase2_single/training_loss.png) -->
|
147 |
+
<img src="./neo_sft_phase2_all_pair/training_loss.png" alt="neo_sft_phase2_all_pair_loss" width="60%">
|