Spaces:
Running
Running
Apply chat template for Atla responses
Browse files- gen_api_answer.py +95 -57
gen_api_answer.py
CHANGED
@@ -12,19 +12,17 @@ from prompts import (
|
|
12 |
PROMETHEUS_PROMPT_WITH_REFERENCE,
|
13 |
ATLA_PROMPT,
|
14 |
ATLA_PROMPT_WITH_REFERENCE,
|
15 |
-
|
16 |
)
|
|
|
17 |
|
18 |
# Initialize clients
|
19 |
anthropic_client = anthropic.Anthropic()
|
20 |
openai_client = OpenAI()
|
21 |
together_client = Together()
|
22 |
hf_api_key = os.getenv("HF_API_KEY")
|
23 |
-
cohere_client = cohere.ClientV2(os.getenv("CO_API_KEY"))
|
24 |
-
|
25 |
-
|
26 |
flow_judge_api_key = os.getenv("FLOW_JUDGE_API_KEY")
|
27 |
-
|
28 |
|
29 |
def get_openai_response(model_name, prompt, system_prompt=JUDGE_SYSTEM_PROMPT, max_tokens=500, temperature=0):
|
30 |
"""Get response from OpenAI API"""
|
@@ -73,7 +71,7 @@ def get_together_response(model_name, prompt, system_prompt=JUDGE_SYSTEM_PROMPT,
|
|
73 |
except Exception as e:
|
74 |
return f"Error with Together model {model_name}: {str(e)}"
|
75 |
|
76 |
-
def get_prometheus_response(model_name, prompt, max_tokens=500, temperature=0.01):
|
77 |
"""Get response from Hugging Face model"""
|
78 |
try:
|
79 |
headers = {
|
@@ -82,8 +80,19 @@ def get_prometheus_response(model_name, prompt, max_tokens=500, temperature=0.01
|
|
82 |
"Content-Type": "application/json"
|
83 |
}
|
84 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
85 |
payload = {
|
86 |
-
"inputs":
|
87 |
"parameters": {
|
88 |
"max_new_tokens": max_tokens,
|
89 |
"return_full_text": False,
|
@@ -100,7 +109,7 @@ def get_prometheus_response(model_name, prompt, max_tokens=500, temperature=0.01
|
|
100 |
except Exception as e:
|
101 |
return f"Error with Hugging Face model {model_name}: {str(e)}"
|
102 |
|
103 |
-
def get_atla_response(model_name, prompt, max_tokens=500, temperature=0.01):
|
104 |
"""Get response from HF endpoint for Atla model"""
|
105 |
try:
|
106 |
headers = {
|
@@ -109,13 +118,25 @@ def get_atla_response(model_name, prompt, max_tokens=500, temperature=0.01):
|
|
109 |
"Content-Type": "application/json"
|
110 |
}
|
111 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
112 |
payload = {
|
113 |
-
"inputs":
|
114 |
"parameters": {
|
115 |
"max_new_tokens": max_tokens,
|
116 |
"return_full_text": False,
|
117 |
"temperature": temperature,
|
118 |
-
"seed": 42
|
|
|
119 |
}
|
120 |
}
|
121 |
|
@@ -128,27 +149,6 @@ def get_atla_response(model_name, prompt, max_tokens=500, temperature=0.01):
|
|
128 |
except Exception as e:
|
129 |
return f"Error with Atla model {model_name}: {str(e)}"
|
130 |
|
131 |
-
def get_cohere_response(model_name, prompt, system_prompt=JUDGE_SYSTEM_PROMPT, max_tokens=500, temperature=0):
|
132 |
-
"""Get response from Cohere API"""
|
133 |
-
try:
|
134 |
-
response = cohere_client.chat(
|
135 |
-
model=model_name,
|
136 |
-
messages=[
|
137 |
-
{"role": "system", "content": system_prompt},
|
138 |
-
{"role": "user", "content": prompt}
|
139 |
-
],
|
140 |
-
max_tokens=max_tokens,
|
141 |
-
temperature=temperature
|
142 |
-
)
|
143 |
-
# Extract the text from the content items
|
144 |
-
content_items = response.message.content
|
145 |
-
if isinstance(content_items, list):
|
146 |
-
# Get the text from the first content item
|
147 |
-
return content_items[0].text
|
148 |
-
return str(content_items) # Fallback if it's not a list
|
149 |
-
except Exception as e:
|
150 |
-
return f"Error with Cohere model {model_name}: {str(e)}"
|
151 |
-
|
152 |
def get_flow_judge_response(model_name, prompt, max_tokens=500, temperature=0.1, top_p=0.95) -> str:
|
153 |
"""Get response from Flow Judge"""
|
154 |
try:
|
@@ -173,6 +173,27 @@ def get_flow_judge_response(model_name, prompt, max_tokens=500, temperature=0.1,
|
|
173 |
except Exception as e:
|
174 |
return f"Error with Flow Judge completions model {model_name}: {str(e)}"
|
175 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
176 |
def get_model_response(
|
177 |
model_name,
|
178 |
model_info,
|
@@ -188,21 +209,22 @@ def get_model_response(
|
|
188 |
api_model = model_info["api_model"]
|
189 |
organization = model_info["organization"]
|
190 |
|
191 |
-
# Determine if model is Prometheus or Atla
|
192 |
is_prometheus = (organization == "Prometheus")
|
193 |
is_atla = (organization == "Atla")
|
194 |
is_flow_judge = (organization == "Flow AI")
|
195 |
-
# For non-Prometheus/Atla models, use the Judge system prompt
|
196 |
system_prompt = None if (is_prometheus or is_atla or is_flow_judge) else JUDGE_SYSTEM_PROMPT
|
197 |
|
198 |
# Select the appropriate base prompt
|
|
|
199 |
if is_atla:
|
200 |
base_prompt = ATLA_PROMPT_WITH_REFERENCE if use_reference else ATLA_PROMPT
|
201 |
elif is_flow_judge:
|
202 |
base_prompt = FLOW_JUDGE_PROMPT
|
203 |
else:
|
204 |
base_prompt = PROMETHEUS_PROMPT_WITH_REFERENCE if use_reference else PROMETHEUS_PROMPT
|
205 |
-
|
206 |
# For non-Prometheus/non-Atla models, replace the specific instruction
|
207 |
if not (is_prometheus or is_atla or is_flow_judge):
|
208 |
base_prompt = base_prompt.replace(
|
@@ -224,6 +246,7 @@ def get_model_response(
|
|
224 |
score4_desc=prompt_data['score4_desc'],
|
225 |
score5_desc=prompt_data['score5_desc']
|
226 |
)
|
|
|
227 |
else:
|
228 |
human_input = f"<user_input>\n{prompt_data['human_input']}\n</user_input>"
|
229 |
ai_response = f"<response>\n{prompt_data['ai_response']}\n</response>"
|
@@ -249,6 +272,7 @@ def get_model_response(
|
|
249 |
EVALUATION_CRITERIA=eval_criteria,
|
250 |
RUBRIC=rubric
|
251 |
)
|
|
|
252 |
except KeyError as e:
|
253 |
return f"Error formatting prompt: Missing required field {str(e)}"
|
254 |
|
@@ -263,11 +287,11 @@ def get_model_response(
|
|
263 |
)
|
264 |
elif organization == "Prometheus":
|
265 |
return get_prometheus_response(
|
266 |
-
api_model, final_prompt, max_tokens, temperature = 0.01
|
267 |
)
|
268 |
elif organization == "Atla":
|
269 |
return get_atla_response(
|
270 |
-
api_model, final_prompt, max_tokens, temperature = 0.01
|
271 |
)
|
272 |
elif organization == "Cohere":
|
273 |
return get_cohere_response(
|
@@ -290,6 +314,10 @@ def parse_model_response(response):
|
|
290 |
# Debug print
|
291 |
print(f"Raw model response: {response}")
|
292 |
|
|
|
|
|
|
|
|
|
293 |
# First try to parse the entire response as JSON
|
294 |
try:
|
295 |
data = json.loads(response)
|
@@ -306,6 +334,16 @@ def parse_model_response(response):
|
|
306 |
except Exception as e:
|
307 |
# Debug print for error case
|
308 |
print(f"Failed to parse response: {str(e)}")
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
309 |
return "Error", f"Failed to parse response: {response}"
|
310 |
|
311 |
def prometheus_parse_model_response(output):
|
@@ -363,6 +401,27 @@ def prometheus_parse_model_response(output):
|
|
363 |
except Exception as e:
|
364 |
print(f"Failed to parse response: {str(e)}")
|
365 |
return "Error", f"Exception during parsing: {str(e)}"
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
366 |
|
367 |
def flow_judge_parse_model_response(output):
|
368 |
try:
|
@@ -386,25 +445,4 @@ def flow_judge_parse_model_response(output):
|
|
386 |
|
387 |
except Exception as e:
|
388 |
print(f"Failed to parse response: {str(e)}")
|
389 |
-
return "Error", f"Exception during parsing: {str(e)}"
|
390 |
-
|
391 |
-
def atla_parse_model_response(output):
|
392 |
-
"""Parse response from ATLA model"""
|
393 |
-
try:
|
394 |
-
print(f"Raw Atla model response: {output}")
|
395 |
-
output = output.strip()
|
396 |
-
|
397 |
-
# Look for the Reasoning and Result sections
|
398 |
-
reasoning_match = re.search(r'\*\*Reasoning:\*\*(.*?)(?=\*\*Result:|$)', output, re.DOTALL)
|
399 |
-
result_match = re.search(r'\*\*Result:\*\*\s*(\d+)', output)
|
400 |
-
|
401 |
-
if reasoning_match and result_match:
|
402 |
-
feedback = reasoning_match.group(1).strip()
|
403 |
-
score = result_match.group(1)
|
404 |
-
return str(score), feedback
|
405 |
-
|
406 |
-
return "Error", f"Failed to parse ATLA response format: {output}"
|
407 |
-
|
408 |
-
except Exception as e:
|
409 |
-
print(f"Failed to parse ATLA response: {str(e)}")
|
410 |
return "Error", f"Exception during parsing: {str(e)}"
|
|
|
12 |
PROMETHEUS_PROMPT_WITH_REFERENCE,
|
13 |
ATLA_PROMPT,
|
14 |
ATLA_PROMPT_WITH_REFERENCE,
|
15 |
+
FLOW_JUDGE_PROMPT
|
16 |
)
|
17 |
+
from transformers import AutoTokenizer
|
18 |
|
19 |
# Initialize clients
|
20 |
anthropic_client = anthropic.Anthropic()
|
21 |
openai_client = OpenAI()
|
22 |
together_client = Together()
|
23 |
hf_api_key = os.getenv("HF_API_KEY")
|
|
|
|
|
|
|
24 |
flow_judge_api_key = os.getenv("FLOW_JUDGE_API_KEY")
|
25 |
+
cohere_client = cohere.ClientV2(os.getenv("CO_API_KEY"))
|
26 |
|
27 |
def get_openai_response(model_name, prompt, system_prompt=JUDGE_SYSTEM_PROMPT, max_tokens=500, temperature=0):
|
28 |
"""Get response from OpenAI API"""
|
|
|
71 |
except Exception as e:
|
72 |
return f"Error with Together model {model_name}: {str(e)}"
|
73 |
|
74 |
+
def get_prometheus_response(model_name, prompt, system_prompt=None, max_tokens=500, temperature=0.01):
|
75 |
"""Get response from Hugging Face model"""
|
76 |
try:
|
77 |
headers = {
|
|
|
80 |
"Content-Type": "application/json"
|
81 |
}
|
82 |
|
83 |
+
# Create messages list for chat template
|
84 |
+
messages = []
|
85 |
+
if system_prompt:
|
86 |
+
messages.append({"role": "system", "content": system_prompt})
|
87 |
+
messages.append({"role": "user", "content": prompt})
|
88 |
+
|
89 |
+
# Apply chat template
|
90 |
+
model_id = "prometheus-eval/prometheus-7b-v2.0"
|
91 |
+
tokenizer = AutoTokenizer.from_pretrained(model_id)
|
92 |
+
formatted_prompt = tokenizer.apply_chat_template(messages, tokenize=False)
|
93 |
+
|
94 |
payload = {
|
95 |
+
"inputs": formatted_prompt,
|
96 |
"parameters": {
|
97 |
"max_new_tokens": max_tokens,
|
98 |
"return_full_text": False,
|
|
|
109 |
except Exception as e:
|
110 |
return f"Error with Hugging Face model {model_name}: {str(e)}"
|
111 |
|
112 |
+
def get_atla_response(model_name, prompt, system_prompt=None, max_tokens=500, temperature=0.01):
|
113 |
"""Get response from HF endpoint for Atla model"""
|
114 |
try:
|
115 |
headers = {
|
|
|
118 |
"Content-Type": "application/json"
|
119 |
}
|
120 |
|
121 |
+
# Create messages list for chat template
|
122 |
+
messages = []
|
123 |
+
if system_prompt:
|
124 |
+
messages.append({"role": "system", "content": system_prompt})
|
125 |
+
messages.append({"role": "user", "content": prompt})
|
126 |
+
|
127 |
+
# Apply chat template
|
128 |
+
model_id = "AtlaAI/Atla-8B-preview" # Update this if using a different model
|
129 |
+
tokenizer = AutoTokenizer.from_pretrained(model_id)
|
130 |
+
formatted_prompt = tokenizer.apply_chat_template(messages, tokenize=False)
|
131 |
+
|
132 |
payload = {
|
133 |
+
"inputs": formatted_prompt,
|
134 |
"parameters": {
|
135 |
"max_new_tokens": max_tokens,
|
136 |
"return_full_text": False,
|
137 |
"temperature": temperature,
|
138 |
+
"seed": 42,
|
139 |
+
"add_generation_prompt": True
|
140 |
}
|
141 |
}
|
142 |
|
|
|
149 |
except Exception as e:
|
150 |
return f"Error with Atla model {model_name}: {str(e)}"
|
151 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
152 |
def get_flow_judge_response(model_name, prompt, max_tokens=500, temperature=0.1, top_p=0.95) -> str:
|
153 |
"""Get response from Flow Judge"""
|
154 |
try:
|
|
|
173 |
except Exception as e:
|
174 |
return f"Error with Flow Judge completions model {model_name}: {str(e)}"
|
175 |
|
176 |
+
def get_cohere_response(model_name, prompt, system_prompt=JUDGE_SYSTEM_PROMPT, max_tokens=500, temperature=0):
|
177 |
+
"""Get response from Cohere API"""
|
178 |
+
try:
|
179 |
+
response = cohere_client.chat(
|
180 |
+
model=model_name,
|
181 |
+
messages=[
|
182 |
+
{"role": "system", "content": system_prompt},
|
183 |
+
{"role": "user", "content": prompt}
|
184 |
+
],
|
185 |
+
max_tokens=max_tokens,
|
186 |
+
temperature=temperature
|
187 |
+
)
|
188 |
+
# Extract the text from the content items
|
189 |
+
content_items = response.message.content
|
190 |
+
if isinstance(content_items, list):
|
191 |
+
# Get the text from the first content item
|
192 |
+
return content_items[0].text
|
193 |
+
return str(content_items) # Fallback if it's not a list
|
194 |
+
except Exception as e:
|
195 |
+
return f"Error with Cohere model {model_name}: {str(e)}"
|
196 |
+
|
197 |
def get_model_response(
|
198 |
model_name,
|
199 |
model_info,
|
|
|
209 |
api_model = model_info["api_model"]
|
210 |
organization = model_info["organization"]
|
211 |
|
212 |
+
# Determine if model is Prometheus or Atla or Flow Judge
|
213 |
is_prometheus = (organization == "Prometheus")
|
214 |
is_atla = (organization == "Atla")
|
215 |
is_flow_judge = (organization == "Flow AI")
|
216 |
+
# For non-Prometheus/Atla models/Flow Judge, use the Judge system prompt
|
217 |
system_prompt = None if (is_prometheus or is_atla or is_flow_judge) else JUDGE_SYSTEM_PROMPT
|
218 |
|
219 |
# Select the appropriate base prompt
|
220 |
+
|
221 |
if is_atla:
|
222 |
base_prompt = ATLA_PROMPT_WITH_REFERENCE if use_reference else ATLA_PROMPT
|
223 |
elif is_flow_judge:
|
224 |
base_prompt = FLOW_JUDGE_PROMPT
|
225 |
else:
|
226 |
base_prompt = PROMETHEUS_PROMPT_WITH_REFERENCE if use_reference else PROMETHEUS_PROMPT
|
227 |
+
|
228 |
# For non-Prometheus/non-Atla models, replace the specific instruction
|
229 |
if not (is_prometheus or is_atla or is_flow_judge):
|
230 |
base_prompt = base_prompt.replace(
|
|
|
246 |
score4_desc=prompt_data['score4_desc'],
|
247 |
score5_desc=prompt_data['score5_desc']
|
248 |
)
|
249 |
+
|
250 |
else:
|
251 |
human_input = f"<user_input>\n{prompt_data['human_input']}\n</user_input>"
|
252 |
ai_response = f"<response>\n{prompt_data['ai_response']}\n</response>"
|
|
|
272 |
EVALUATION_CRITERIA=eval_criteria,
|
273 |
RUBRIC=rubric
|
274 |
)
|
275 |
+
|
276 |
except KeyError as e:
|
277 |
return f"Error formatting prompt: Missing required field {str(e)}"
|
278 |
|
|
|
287 |
)
|
288 |
elif organization == "Prometheus":
|
289 |
return get_prometheus_response(
|
290 |
+
api_model, final_prompt, system_prompt, max_tokens, temperature = 0.01
|
291 |
)
|
292 |
elif organization == "Atla":
|
293 |
return get_atla_response(
|
294 |
+
api_model, final_prompt, system_prompt, max_tokens, temperature = 0.01
|
295 |
)
|
296 |
elif organization == "Cohere":
|
297 |
return get_cohere_response(
|
|
|
314 |
# Debug print
|
315 |
print(f"Raw model response: {response}")
|
316 |
|
317 |
+
# If response is already a dictionary, use it directly
|
318 |
+
if isinstance(response, dict):
|
319 |
+
return str(response.get("result", "N/A")), response.get("feedback", "N/A")
|
320 |
+
|
321 |
# First try to parse the entire response as JSON
|
322 |
try:
|
323 |
data = json.loads(response)
|
|
|
334 |
except Exception as e:
|
335 |
# Debug print for error case
|
336 |
print(f"Failed to parse response: {str(e)}")
|
337 |
+
|
338 |
+
# If the error message itself contains valid JSON, try to parse that
|
339 |
+
try:
|
340 |
+
error_json_match = re.search(r"{.*}", str(e), re.DOTALL)
|
341 |
+
if error_json_match:
|
342 |
+
data = json.loads(error_json_match.group(0))
|
343 |
+
return str(data.get("result", "N/A")), data.get("feedback", "N/A")
|
344 |
+
except:
|
345 |
+
pass
|
346 |
+
|
347 |
return "Error", f"Failed to parse response: {response}"
|
348 |
|
349 |
def prometheus_parse_model_response(output):
|
|
|
401 |
except Exception as e:
|
402 |
print(f"Failed to parse response: {str(e)}")
|
403 |
return "Error", f"Exception during parsing: {str(e)}"
|
404 |
+
|
405 |
+
def atla_parse_model_response(output):
|
406 |
+
"""Parse response from ATLA model"""
|
407 |
+
try:
|
408 |
+
print(f"Raw Atla model response: {output}")
|
409 |
+
output = output.strip()
|
410 |
+
|
411 |
+
# Look for the Reasoning and Result sections
|
412 |
+
reasoning_match = re.search(r'\*\*Reasoning:\*\*(.*?)(?=\*\*Result:|$)', output, re.DOTALL)
|
413 |
+
result_match = re.search(r'\*\*Result:\*\*\s*(\d+)', output)
|
414 |
+
|
415 |
+
if reasoning_match and result_match:
|
416 |
+
feedback = reasoning_match.group(1).strip()
|
417 |
+
score = result_match.group(1)
|
418 |
+
return str(score), feedback
|
419 |
+
|
420 |
+
return "Error", f"Failed to parse ATLA response format: {output}"
|
421 |
+
|
422 |
+
except Exception as e:
|
423 |
+
print(f"Failed to parse ATLA response: {str(e)}")
|
424 |
+
return "Error", f"Exception during parsing: {str(e)}"
|
425 |
|
426 |
def flow_judge_parse_model_response(output):
|
427 |
try:
|
|
|
445 |
|
446 |
except Exception as e:
|
447 |
print(f"Failed to parse response: {str(e)}")
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
448 |
return "Error", f"Exception during parsing: {str(e)}"
|