Update README.md
Browse files
README.md
CHANGED
@@ -1,3 +1,107 @@
|
|
1 |
-
---
|
2 |
-
license: llama3
|
3 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
---
|
2 |
+
license: llama3
|
3 |
+
datasets:
|
4 |
+
- REILX/neo_sft_phase2_single
|
5 |
+
- REILX/neo_sft_phase2_multi
|
6 |
+
- REILX/neo_sft_phase2_conversations
|
7 |
+
language:
|
8 |
+
- en
|
9 |
+
- zh
|
10 |
+
---
|
11 |
+
|
12 |
+
### 数据集
|
13 |
+
|
14 |
+
以 m-a-p/neo_sft_phase2 数据集为基石,精心构建了三个子数据集,分别如下:
|
15 |
+
|
16 |
+
1. REILX/neo_sft_phase2_conversations
|
17 |
+
2. REILX/neo_sft_phase2_multi
|
18 |
+
3. REILX/neo_sft_phase2_single
|
19 |
+
|
20 |
+
### 数据集构建规则
|
21 |
+
|
22 |
+
**REILX/neo_sft_phase2_conversations**
|
23 |
+
|
24 |
+
* **方法:** 将每轮对话视作独立的问答对,并巧妙融入上下文信息构建样本。
|
25 |
+
* **具体步骤:**
|
26 |
+
1. 针对每个“conversation”,逐一遍历其对话轮次。
|
27 |
+
2. 将当前“human”轮次的“value”与之前所有轮次的对话内容拼接,构成完整的“instruction”。
|
28 |
+
3. 将当前“gpt”轮次的“value”作为最终的“output”。
|
29 |
+
4. “input”可为空白,亦可注入适当的提示信息。
|
30 |
+
|
31 |
+
**REILX/neo_sft_phase2_multi**
|
32 |
+
|
33 |
+
* **方法:** 将每轮对话视作独立的问答对,并巧妙利用上下文信息构建样本。
|
34 |
+
* **具体步骤:**
|
35 |
+
1. 针对每个“conversation”,逐一遍历其对话轮次。
|
36 |
+
2. 将每个“conversation”中所有“human”的“value”拼接,构成完整的“instruction”。
|
37 |
+
3. 将每个“conversation”中所有“gpt”的“value”拼接,构成最终的“output”。
|
38 |
+
4. “input”可为空白,亦可注入适当的提示信息。
|
39 |
+
|
40 |
+
**REILX/neo_sft_phase2_single**
|
41 |
+
|
42 |
+
* **具体步骤:**
|
43 |
+
1. 针对每个“conversation”,逐一遍历其对话轮次。
|
44 |
+
2. 仅保留包含一轮对话的“conversation”,舍弃多轮对话数据。
|
45 |
+
3. 将该“conversation”的“human”的“value”作为“instruction”。
|
46 |
+
4. 将该“conversation”的“gpt”的“value”作为“output”。
|
47 |
+
5. “input”可为空白,亦可注入适当的提示信息。
|
48 |
+
|
49 |
+
### 训练参数
|
50 |
+
REILX/neo_sft_phase2_conversations</br>
|
51 |
+
The following hyperparameters were used during training:
|
52 |
+
- learning_rate: 2e-05
|
53 |
+
- train_batch_size: 1
|
54 |
+
- eval_batch_size: 8
|
55 |
+
- cutoff_len: 8192
|
56 |
+
- seed: 42
|
57 |
+
- distributed_type: multi-GPU
|
58 |
+
- num_devices: 8
|
59 |
+
- gradient_accumulation_steps: 8
|
60 |
+
- total_train_batch_size: 64
|
61 |
+
- total_eval_batch_size: 64
|
62 |
+
- optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
|
63 |
+
- lr_scheduler_type: cosine
|
64 |
+
- lr_scheduler_warmup_ratio: 0.1
|
65 |
+
- num_epochs: 5.0
|
66 |
+
|
67 |
+
REILX/neo_sft_phase2_multi</br>
|
68 |
+
- learning_rate: 2e-05
|
69 |
+
- train_batch_size: 1
|
70 |
+
- eval_batch_size: 8
|
71 |
+
- cutoff_len: 8192
|
72 |
+
- seed: 42
|
73 |
+
- distributed_type: multi-GPU
|
74 |
+
- num_devices: 8
|
75 |
+
- gradient_accumulation_steps: 8
|
76 |
+
- total_train_batch_size: 64
|
77 |
+
- total_eval_batch_size: 64
|
78 |
+
- optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
|
79 |
+
- lr_scheduler_type: cosine
|
80 |
+
- lr_scheduler_warmup_ratio: 0.1
|
81 |
+
- num_epochs: 5.0
|
82 |
+
|
83 |
+
REILX/neo_sft_phase2_single</br>
|
84 |
+
- learning_rate: 2e-05
|
85 |
+
- train_batch_size: 1
|
86 |
+
- eval_batch_size: 8
|
87 |
+
- cutoff_len: 4096
|
88 |
+
- seed: 42
|
89 |
+
- distributed_type: multi-GPU
|
90 |
+
- num_devices: 8
|
91 |
+
- gradient_accumulation_steps: 8
|
92 |
+
- total_train_batch_size: 64
|
93 |
+
- total_eval_batch_size: 64
|
94 |
+
- optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
|
95 |
+
- lr_scheduler_type: cosine
|
96 |
+
- lr_scheduler_warmup_ratio: 0.1
|
97 |
+
- num_epochs: 5.0
|
98 |
+
|
99 |
+
### 损失图
|
100 |
+
REILX/neo_sft_phase2_conversations</br>
|
101 |
+
<img src="./neo_sft_phase2_conversations/training_loss.png" alt="neo_sft_phase2_conversations_loss" width="60%">
|
102 |
+
|
103 |
+
REILX/neo_sft_phase2_multi</br>
|
104 |
+
<img src="./neo_sft_phase2_multi/training_loss.png" alt="neo_sft_phase2_multi_loss" width="60%">
|
105 |
+
|
106 |
+
REILX/neo_sft_phase2_single</br>
|
107 |
+
<img src="./neo_sft_phase2_single/training_loss.png" alt="neo_sft_phase2_single_loss" width="60%">
|