Update README.md
Browse files
README.md
CHANGED
@@ -28,47 +28,8 @@ Eurus-RM-7B is trained on a mixture of [UltraInteract](https://huggingface.co/da
|
|
28 |
|
29 |
## Usage
|
30 |
```python
|
31 |
-
from transformers import
|
32 |
-
import torch.nn as nn
|
33 |
-
import torch
|
34 |
-
from typing import Optional, List
|
35 |
|
36 |
-
class EurusRewardModel(PreTrainedModel):
|
37 |
-
config_class = MistralConfig
|
38 |
-
def __init__(self, config):
|
39 |
-
super().__init__(config)
|
40 |
-
self.model = MistralModel(config)
|
41 |
-
self.regression_head = nn.Linear(self.config.hidden_size, 1, bias=False)
|
42 |
-
|
43 |
-
def forward( # args are the same as LlamaForCausalLM
|
44 |
-
self,
|
45 |
-
input_ids: torch.LongTensor = None,
|
46 |
-
attention_mask: Optional[torch.Tensor] = None,
|
47 |
-
position_ids: Optional[torch.LongTensor] = None,
|
48 |
-
past_key_values: Optional[List[torch.FloatTensor]] = None,
|
49 |
-
inputs_embeds: Optional[torch.FloatTensor] = None,
|
50 |
-
labels: Optional[torch.LongTensor] = None,
|
51 |
-
use_cache: Optional[bool] = None,
|
52 |
-
output_attentions: Optional[bool] = None,
|
53 |
-
output_hidden_states: Optional[bool] = None,
|
54 |
-
return_dict: Optional[bool] = None,
|
55 |
-
):
|
56 |
-
|
57 |
-
transformer_outputs = self.model(
|
58 |
-
input_ids,
|
59 |
-
attention_mask=attention_mask,
|
60 |
-
position_ids=position_ids,
|
61 |
-
past_key_values=past_key_values,
|
62 |
-
inputs_embeds=inputs_embeds,
|
63 |
-
)
|
64 |
-
|
65 |
-
hidden_states = transformer_outputs[0]
|
66 |
-
rewards = self.regression_head(hidden_states).squeeze(-1)
|
67 |
-
|
68 |
-
ends = attention_mask.cumsum(dim=1).argmax(dim=1).view(-1,1)
|
69 |
-
rewards = torch.gather(rewards, 1, ends)
|
70 |
-
|
71 |
-
return rewards
|
72 |
|
73 |
def test(model_path):
|
74 |
dataset = [ # cases in webgpt; we use the same template as Mistral-Instruct-v0.2
|
@@ -78,7 +39,7 @@ def test(model_path):
|
|
78 |
|
79 |
|
80 |
tokenizer = AutoTokenizer.from_pretrained(model_path)
|
81 |
-
model =
|
82 |
|
83 |
for example in dataset:
|
84 |
inputs = tokenizer(example["chosen"], return_tensors="pt")
|
|
|
28 |
|
29 |
## Usage
|
30 |
```python
|
31 |
+
from transformers import AutoTokenizer, AutoModel
|
|
|
|
|
|
|
32 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
33 |
|
34 |
def test(model_path):
|
35 |
dataset = [ # cases in webgpt; we use the same template as Mistral-Instruct-v0.2
|
|
|
39 |
|
40 |
|
41 |
tokenizer = AutoTokenizer.from_pretrained(model_path)
|
42 |
+
model = AutoModel.from_pretrained(model_path, trust_remote_code=True)
|
43 |
|
44 |
for example in dataset:
|
45 |
inputs = tokenizer(example["chosen"], return_tensors="pt")
|