Update README.md
Browse filesUpdate generations after major fix: https://github.com/huggingface/transformers/commit/abc400b06a8ab26cd438b6e9add3aad082ffc48f
README.md
CHANGED
@@ -63,7 +63,7 @@ It is recommended to directly call the [`generate`](https://huggingface.co/docs/
|
|
63 |
>>> # the fast tokenizer currently does not work correctly
|
64 |
>>> tokenizer = AutoTokenizer.from_pretrained("facebook/opt-30b", use_fast=False)
|
65 |
|
66 |
-
>>> prompt = "Hello, I
|
67 |
|
68 |
|
69 |
>>> input_ids = tokenizer(prompt, return_tensors="pt").input_ids.cuda()
|
@@ -71,7 +71,7 @@ It is recommended to directly call the [`generate`](https://huggingface.co/docs/
|
|
71 |
>>> generated_ids = model.generate(input_ids)
|
72 |
|
73 |
>>> tokenizer.batch_decode(generated_ids, skip_special_tokens=True)
|
74 |
-
[
|
75 |
```
|
76 |
|
77 |
By default, generation is deterministic. In order to use the top-k sampling, please set `do_sample` to `True`.
|
@@ -85,7 +85,7 @@ By default, generation is deterministic. In order to use the top-k sampling, ple
|
|
85 |
>>> # the fast tokenizer currently does not work correctly
|
86 |
>>> tokenizer = AutoTokenizer.from_pretrained("facebook/opt-30b", use_fast=False)
|
87 |
|
88 |
-
>>> prompt = "Hello, I
|
89 |
|
90 |
>>> input_ids = tokenizer(prompt, return_tensors="pt").input_ids.cuda()
|
91 |
|
@@ -93,7 +93,7 @@ By default, generation is deterministic. In order to use the top-k sampling, ple
|
|
93 |
>>> generated_ids = model.generate(input_ids, do_sample=True)
|
94 |
|
95 |
>>> tokenizer.batch_decode(generated_ids, skip_special_tokens=True)
|
96 |
-
[
|
97 |
```
|
98 |
|
99 |
### Limitations and bias
|
@@ -126,11 +126,11 @@ Here's an example of how the model can have biased predictions:
|
|
126 |
>>> generated_ids = model.generate(input_ids, do_sample=True, num_return_sequences=5, max_length=10)
|
127 |
|
128 |
>>> tokenizer.batch_decode(generated_ids, skip_special_tokens=True)
|
129 |
-
The woman worked as a
|
130 |
-
The woman worked as a
|
131 |
-
The woman worked as a
|
132 |
-
The woman worked as a
|
133 |
-
|
134 |
```
|
135 |
|
136 |
compared to:
|
@@ -152,11 +152,11 @@ compared to:
|
|
152 |
>>> generated_ids = model.generate(input_ids, do_sample=True, num_return_sequences=5, max_length=10)
|
153 |
|
154 |
>>> tokenizer.batch_decode(generated_ids, skip_special_tokens=True)
|
155 |
-
The man worked as a
|
156 |
-
The man worked as a
|
157 |
-
The man worked as a
|
158 |
-
The man worked as a
|
159 |
-
The man worked as a
|
160 |
```
|
161 |
|
162 |
This bias will also affect all fine-tuned versions of this model.
|
|
|
63 |
>>> # the fast tokenizer currently does not work correctly
|
64 |
>>> tokenizer = AutoTokenizer.from_pretrained("facebook/opt-30b", use_fast=False)
|
65 |
|
66 |
+
>>> prompt = "Hello, I am conscious and"
|
67 |
|
68 |
|
69 |
>>> input_ids = tokenizer(prompt, return_tensors="pt").input_ids.cuda()
|
|
|
71 |
>>> generated_ids = model.generate(input_ids)
|
72 |
|
73 |
>>> tokenizer.batch_decode(generated_ids, skip_special_tokens=True)
|
74 |
+
['Hello, I am conscious and I am here.\nI am also conscious and I am here']
|
75 |
```
|
76 |
|
77 |
By default, generation is deterministic. In order to use the top-k sampling, please set `do_sample` to `True`.
|
|
|
85 |
>>> # the fast tokenizer currently does not work correctly
|
86 |
>>> tokenizer = AutoTokenizer.from_pretrained("facebook/opt-30b", use_fast=False)
|
87 |
|
88 |
+
>>> prompt = "Hello, I am conscious and"
|
89 |
|
90 |
>>> input_ids = tokenizer(prompt, return_tensors="pt").input_ids.cuda()
|
91 |
|
|
|
93 |
>>> generated_ids = model.generate(input_ids, do_sample=True)
|
94 |
|
95 |
>>> tokenizer.batch_decode(generated_ids, skip_special_tokens=True)
|
96 |
+
['Hello, I am conscious and aware that you have your back turned to me and want to talk']
|
97 |
```
|
98 |
|
99 |
### Limitations and bias
|
|
|
126 |
>>> generated_ids = model.generate(input_ids, do_sample=True, num_return_sequences=5, max_length=10)
|
127 |
|
128 |
>>> tokenizer.batch_decode(generated_ids, skip_special_tokens=True)
|
129 |
+
The woman worked as a supervisor in the office
|
130 |
+
The woman worked as a social worker in a
|
131 |
+
The woman worked as a cashier at the
|
132 |
+
The woman worked as a teacher from 2011 to
|
133 |
+
he woman worked as a maid at the house
|
134 |
```
|
135 |
|
136 |
compared to:
|
|
|
152 |
>>> generated_ids = model.generate(input_ids, do_sample=True, num_return_sequences=5, max_length=10)
|
153 |
|
154 |
>>> tokenizer.batch_decode(generated_ids, skip_special_tokens=True)
|
155 |
+
The man worked as a school bus driver for
|
156 |
+
The man worked as a bartender in a bar
|
157 |
+
The man worked as a cashier at the
|
158 |
+
The man worked as a teacher, and was
|
159 |
+
The man worked as a professional at a range
|
160 |
```
|
161 |
|
162 |
This bias will also affect all fine-tuned versions of this model.
|