liujch1998 commited on
Commit
0f3b775
·
1 Parent(s): 9302ce1

Update README.md

Browse files
Files changed (1) hide show
  1. README.md +46 -0
README.md CHANGED
@@ -155,6 +155,52 @@ Users (both direct and downstream) should be made aware of the risks, biases and
155
  Use the code below to get started with the model.
156
 
157
  ```python
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
158
  ```
159
 
160
  You may also refer to <https://huggingface.co/spaces/liujch1998/crystal/blob/main/app.py#L10-L86> for implementation.
 
155
  Use the code below to get started with the model.
156
 
157
  ```python
158
+ from transformers import T5Tokenizer, T5ForConditionalGeneration
159
+
160
+ tokenizer = AutoTokenizer.from_pretrained('liujch1998/crystal-11b')
161
+ model = AutoModelForSeq2SeqLM.from_pretrained('liujch1998/crystal-11b')
162
+ model.eval()
163
+
164
+ max_question_len, max_knowledge_len, max_answer_len = 128, 32, 2
165
+ k = 1 # number of knowledge statements to generate
166
+ top_p = 0.0001
167
+
168
+ question = 'If the mass of an object gets bigger what will happen to the amount of matter contained within it? \\n (A) gets bigger (B) gets smaller'
169
+ choices = ['A', 'B']
170
+
171
+ choices_ids = tokenizer(choices, return_tensors='pt', padding='max_length', truncation='longest_first', max_length=max_answer_len).input_ids # (C, AL)
172
+
173
+ prompt = question + ' \\n Knowledge: '
174
+ prompt_tok = tokenizer(prompt, return_tensors='pt', padding='max_length', truncation='longest_first', max_length=max_question_len) # (1, QL)
175
+ knowledges_ids = self.model.generate(
176
+ input_ids=prompt_tok.input_ids,
177
+ attention_mask=prompt_tok.attention_mask,
178
+ max_length=max_knowledge_len + 1,
179
+ min_length=3,
180
+ do_sample=True,
181
+ num_return_sequences=k,
182
+ top_p=top_p,
183
+ ) # (K, KL); begins with 0 ([BOS]); ends with 1 ([EOS])
184
+ knowledges_ids = knowledges_ids[:, 1:].contiguous() # no beginning; ends with 1 ([EOS])
185
+ knowledges = tokenizer.batch_decode(knowledges_ids, skip_special_tokens=True, clean_up_tokenization_spaces=True)
186
+
187
+ prompts = [question + (f' \\n Knowledge: {knowledge} \\n Answer: ' if knowledge != '' else ' \\n Answer:') for knowledge in knowledges]
188
+ prompts_tok = self.tokenizer(prompts, return_tensors='pt', padding='max_length', truncation='longest_first', max_length=max_question_len + max_knowledge_len) # (K, QL+KL)
189
+ output = model(
190
+ input_ids=prompts_tok.input_ids,
191
+ attention_mask=prompts_tok.attention_mask,
192
+ labels=choices_ids[0].unsqueeze(0).repeat(len(knowledges), 1),
193
+ )
194
+ logitsss = output.logits # (K, AL, V)
195
+ logitss = logitsss[:, 0, :] # (K, V)
196
+ choice_ids = choices_ids[:, 0] # (C)
197
+ answer_logitss = logitss.gather(dim=1, index=choice_ids.unsqueeze(0).expand(len(knowledges), -1)) # (K, C)
198
+ answer_probss = answer_logitss.softmax(dim=1) # (K, C)
199
+ answer_probs = answer_probss.max(dim=0).values # (C)
200
+ pred = answer_probs.argmax(dim=0).item()
201
+ pred = choices[pred]
202
+
203
+ print(f'Question: {question}\nKnowledge: {knowledges[0]}\nAnswer: {pred}')
204
  ```
205
 
206
  You may also refer to <https://huggingface.co/spaces/liujch1998/crystal/blob/main/app.py#L10-L86> for implementation.