Upload app.py
Browse files
app.py
CHANGED
@@ -1,18 +1,28 @@
|
|
1 |
# app.py โโ Math-solution classifier for HF Spaces
|
2 |
-
#
|
|
|
|
|
3 |
|
4 |
import os
|
|
|
5 |
import logging
|
|
|
6 |
|
7 |
import gradio as gr
|
8 |
import torch
|
9 |
-
from transformers import
|
|
|
|
|
|
|
10 |
|
11 |
-
#
|
12 |
try:
|
13 |
-
from peft import
|
|
|
|
|
|
|
14 |
PEFT_AVAILABLE = True
|
15 |
-
except ImportError:
|
16 |
PEFT_AVAILABLE = False
|
17 |
|
18 |
# โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ
|
@@ -21,55 +31,75 @@ except ImportError:
|
|
21 |
logging.basicConfig(level=logging.INFO)
|
22 |
logger = logging.getLogger(__name__)
|
23 |
|
24 |
-
ADAPTER_PATH
|
25 |
FALLBACK_MODEL = "distilbert-base-uncased"
|
26 |
-
LABELS
|
27 |
-
|
28 |
-
|
29 |
|
30 |
device = "cuda" if torch.cuda.is_available() else "cpu"
|
31 |
-
|
|
|
32 |
tokenizer = None
|
|
|
33 |
|
34 |
# โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ
|
35 |
-
#
|
36 |
# โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ
|
37 |
def load_model():
|
38 |
-
"""
|
39 |
-
global model, tokenizer
|
|
|
|
|
40 |
|
41 |
if PEFT_AVAILABLE and os.path.isdir(ADAPTER_PATH):
|
42 |
-
logger.info(f"
|
43 |
-
|
44 |
-
|
45 |
-
|
46 |
-
|
47 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
48 |
tokenizer = AutoTokenizer.from_pretrained(ADAPTER_PATH)
|
|
|
49 |
else:
|
50 |
-
logger.warning("
|
51 |
tokenizer = AutoTokenizer.from_pretrained(FALLBACK_MODEL)
|
52 |
-
model
|
53 |
FALLBACK_MODEL,
|
54 |
num_labels=3,
|
55 |
ignore_mismatched_sizes=True,
|
|
|
56 |
)
|
|
|
57 |
|
|
|
58 |
if tokenizer.pad_token is None:
|
59 |
tokenizer.pad_token = tokenizer.eos_token or tokenizer.sep_token
|
60 |
|
61 |
model.to(device)
|
62 |
model.eval()
|
63 |
-
logger.info("Model
|
64 |
|
65 |
# โโโโโโโโโโโโโโโโโโโโโโโโโ๏ฟฝ๏ฟฝ๏ฟฝโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ
|
66 |
-
# Inference
|
67 |
# โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ
|
68 |
-
def
|
69 |
-
"""Return (label, confidence, placeholder-explanation)."""
|
70 |
-
if not question.strip() or not solution.strip():
|
71 |
-
return "Please provide both question and solution.", "", ""
|
72 |
-
|
73 |
text = f"Question: {question}\n\nSolution:\n{solution}"
|
74 |
inputs = tokenizer(
|
75 |
text,
|
@@ -81,11 +111,64 @@ def classify(question: str, solution: str):
|
|
81 |
|
82 |
with torch.no_grad():
|
83 |
logits = model(**inputs).logits
|
84 |
-
probs
|
85 |
-
pred
|
86 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
87 |
|
88 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
89 |
|
90 |
# โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ
|
91 |
# Build Gradio UI
|
@@ -95,27 +178,27 @@ load_model()
|
|
95 |
with gr.Blocks(title="Math Solution Classifier") as demo:
|
96 |
gr.Markdown("# ๐งฎ Math Solution Classifier")
|
97 |
gr.Markdown(
|
98 |
-
"Classify a student
|
99 |
"or **computationally flawed**."
|
100 |
)
|
101 |
|
102 |
with gr.Row():
|
103 |
with gr.Column():
|
104 |
-
q_in
|
105 |
-
s_in
|
106 |
-
btn
|
107 |
with gr.Column():
|
108 |
verdict = gr.Textbox(label="Verdict", interactive=False)
|
109 |
-
conf
|
110 |
-
|
111 |
|
112 |
-
btn.click(classify, [q_in, s_in], [verdict, conf,
|
113 |
|
114 |
gr.Examples(
|
115 |
[
|
116 |
["Solve for x: 2x + 5 = 13", "2x + 5 = 13\n2x = 8\nx = 4"],
|
117 |
["Find the derivative of f(x)=xยฒ", "f'(x)=2x+1"],
|
118 |
-
["What is 15
|
119 |
],
|
120 |
inputs=[q_in, s_in],
|
121 |
)
|
|
|
1 |
# app.py โโ Math-solution classifier for HF Spaces
|
2 |
+
# Compatible with both LoRA-classification and LoRA-causal-LM adapters
|
3 |
+
# Requirements (pin in requirements.txt):
|
4 |
+
# gradio torch transformers peft accelerate spaces
|
5 |
|
6 |
import os
|
7 |
+
import json
|
8 |
import logging
|
9 |
+
from typing import Tuple
|
10 |
|
11 |
import gradio as gr
|
12 |
import torch
|
13 |
+
from transformers import (
|
14 |
+
AutoTokenizer,
|
15 |
+
AutoModelForSequenceClassification,
|
16 |
+
)
|
17 |
|
18 |
+
# PEFT imports (optional)
|
19 |
try:
|
20 |
+
from peft.auto import (
|
21 |
+
AutoPeftModelForSequenceClassification,
|
22 |
+
AutoPeftModelForCausalLM,
|
23 |
+
)
|
24 |
PEFT_AVAILABLE = True
|
25 |
+
except ImportError: # PEFT not installed
|
26 |
PEFT_AVAILABLE = False
|
27 |
|
28 |
# โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ
|
|
|
31 |
logging.basicConfig(level=logging.INFO)
|
32 |
logger = logging.getLogger(__name__)
|
33 |
|
34 |
+
ADAPTER_PATH = os.getenv("ADAPTER_PATH", "./lora_adapter") # local dir or Hub ID
|
35 |
FALLBACK_MODEL = "distilbert-base-uncased"
|
36 |
+
LABELS = {0: "โ
Correct",
|
37 |
+
1: "๐ค Conceptual Error",
|
38 |
+
2: "๐ข Computational Error"}
|
39 |
|
40 |
device = "cuda" if torch.cuda.is_available() else "cpu"
|
41 |
+
|
42 |
+
model = None
|
43 |
tokenizer = None
|
44 |
+
model_ty = None # "classification" | "causal_lm" | "baseline"
|
45 |
|
46 |
# โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ
|
47 |
+
# Model loader
|
48 |
# โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ
|
49 |
def load_model():
|
50 |
+
"""Try adapter as classifier โ causal-LM โ plain baseline."""
|
51 |
+
global model, tokenizer, model_ty
|
52 |
+
|
53 |
+
dtype = torch.float16 if device == "cuda" else torch.float32
|
54 |
|
55 |
if PEFT_AVAILABLE and os.path.isdir(ADAPTER_PATH):
|
56 |
+
logger.info(f"Found adapter at {ADAPTER_PATH}")
|
57 |
+
|
58 |
+
# 1) Try sequence-classification adapter
|
59 |
+
try:
|
60 |
+
model = AutoPeftModelForSequenceClassification.from_pretrained(
|
61 |
+
ADAPTER_PATH,
|
62 |
+
torch_dtype=dtype,
|
63 |
+
device_map="auto" if device == "cuda" else None,
|
64 |
+
)
|
65 |
+
model_ty = "classification"
|
66 |
+
logger.info("Loaded adapter as sequence-classifier")
|
67 |
+
except ValueError:
|
68 |
+
# 2) Fall back to causal-LM adapter
|
69 |
+
logger.info("Adapter is not a classifier โ trying causal-LM")
|
70 |
+
model = AutoPeftModelForCausalLM.from_pretrained(
|
71 |
+
ADAPTER_PATH,
|
72 |
+
torch_dtype=dtype,
|
73 |
+
device_map="auto" if device == "cuda" else None,
|
74 |
+
)
|
75 |
+
model_ty = "causal_lm"
|
76 |
+
logger.info("Loaded adapter as causal-LM")
|
77 |
+
|
78 |
tokenizer = AutoTokenizer.from_pretrained(ADAPTER_PATH)
|
79 |
+
|
80 |
else:
|
81 |
+
logger.warning("No adapter found โ using baseline DistilBERT classifier")
|
82 |
tokenizer = AutoTokenizer.from_pretrained(FALLBACK_MODEL)
|
83 |
+
model = AutoModelForSequenceClassification.from_pretrained(
|
84 |
FALLBACK_MODEL,
|
85 |
num_labels=3,
|
86 |
ignore_mismatched_sizes=True,
|
87 |
+
torch_dtype=dtype,
|
88 |
)
|
89 |
+
model_ty = "baseline"
|
90 |
|
91 |
+
# Make sure we have a pad token
|
92 |
if tokenizer.pad_token is None:
|
93 |
tokenizer.pad_token = tokenizer.eos_token or tokenizer.sep_token
|
94 |
|
95 |
model.to(device)
|
96 |
model.eval()
|
97 |
+
logger.info(f"Model ready on {device} as {model_ty}")
|
98 |
|
99 |
# โโโโโโโโโโโโโโโโโโโโโโโโโ๏ฟฝ๏ฟฝ๏ฟฝโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ
|
100 |
+
# Inference helpers
|
101 |
# โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ
|
102 |
+
def _classify_logits(question: str, solution: str) -> Tuple[str, str, str]:
|
|
|
|
|
|
|
|
|
103 |
text = f"Question: {question}\n\nSolution:\n{solution}"
|
104 |
inputs = tokenizer(
|
105 |
text,
|
|
|
111 |
|
112 |
with torch.no_grad():
|
113 |
logits = model(**inputs).logits
|
114 |
+
probs = torch.softmax(logits, dim=-1)[0]
|
115 |
+
pred = int(torch.argmax(probs))
|
116 |
+
conf = f"{probs[pred].item():.3f}"
|
117 |
+
|
118 |
+
return LABELS.get(pred, "Unknown"), conf, "โ"
|
119 |
+
|
120 |
+
def _classify_generate(question: str, solution: str) -> Tuple[str, str, str]:
|
121 |
+
# Prompt must match the format you used in tuning
|
122 |
+
prompt = (
|
123 |
+
"You are a mathematics tutor.\n"
|
124 |
+
"You are given a math word problem and a student's solution. Decide whether the solution is correct.\n\n"
|
125 |
+
"- Correct = all reasoning and calculations are correct.\n"
|
126 |
+
"- Conceptual Error = reasoning is wrong.\n"
|
127 |
+
"- Computational Error = reasoning okay but arithmetic off.\n\n"
|
128 |
+
"Reply with ONLY one of these JSON lines:\n"
|
129 |
+
'{"verdict": "correct"}\n'
|
130 |
+
'{"verdict": "conceptual"}\n'
|
131 |
+
'{"verdict": "computational"}\n\n"
|
132 |
+
f"Question: {question}\n\nSolution:\n{solution}\n\nAnswer:"
|
133 |
+
)
|
134 |
+
|
135 |
+
inputs = tokenizer(prompt, return_tensors="pt").to(device)
|
136 |
+
with torch.no_grad():
|
137 |
+
out_ids = model.generate(
|
138 |
+
**inputs,
|
139 |
+
max_new_tokens=32,
|
140 |
+
pad_token_id=tokenizer.eos_token_id,
|
141 |
+
)
|
142 |
+
generated = tokenizer.decode(out_ids[0][inputs["input_ids"].shape[1]:],
|
143 |
+
skip_special_tokens=True).strip()
|
144 |
+
|
145 |
+
# Try to parse last JSON line
|
146 |
+
verdict = "Unparsed"
|
147 |
+
try:
|
148 |
+
line = generated.splitlines()[-1]
|
149 |
+
data = json.loads(line)
|
150 |
+
v = data.get("verdict", "").lower()
|
151 |
+
if v.startswith("corr"):
|
152 |
+
verdict = LABELS[0]
|
153 |
+
elif v.startswith("conc"):
|
154 |
+
verdict = LABELS[1]
|
155 |
+
elif v.startswith("comp"):
|
156 |
+
verdict = LABELS[2]
|
157 |
+
except Exception:
|
158 |
+
pass
|
159 |
+
|
160 |
+
return verdict, "", generated
|
161 |
|
162 |
+
def classify(question: str, solution: str):
|
163 |
+
if not question.strip() or not solution.strip():
|
164 |
+
return "Please enter both fields.", "", ""
|
165 |
+
|
166 |
+
if model_ty in ("classification", "baseline"):
|
167 |
+
return _classify_logits(question, solution)
|
168 |
+
elif model_ty == "causal_lm":
|
169 |
+
return _classify_generate(question, solution)
|
170 |
+
else:
|
171 |
+
return "Model not loaded.", "", ""
|
172 |
|
173 |
# โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ
|
174 |
# Build Gradio UI
|
|
|
178 |
with gr.Blocks(title="Math Solution Classifier") as demo:
|
179 |
gr.Markdown("# ๐งฎ Math Solution Classifier")
|
180 |
gr.Markdown(
|
181 |
+
"Classify a student's math solution as **correct**, **conceptually flawed**, "
|
182 |
"or **computationally flawed**."
|
183 |
)
|
184 |
|
185 |
with gr.Row():
|
186 |
with gr.Column():
|
187 |
+
q_in = gr.Textbox(label="Math Question", lines=3)
|
188 |
+
s_in = gr.Textbox(label="Proposed Solution", lines=6)
|
189 |
+
btn = gr.Button("Classify", variant="primary")
|
190 |
with gr.Column():
|
191 |
verdict = gr.Textbox(label="Verdict", interactive=False)
|
192 |
+
conf = gr.Textbox(label="Confidence", interactive=False)
|
193 |
+
raw = gr.Textbox(label="Model Output", interactive=False)
|
194 |
|
195 |
+
btn.click(classify, [q_in, s_in], [verdict, conf, raw])
|
196 |
|
197 |
gr.Examples(
|
198 |
[
|
199 |
["Solve for x: 2x + 5 = 13", "2x + 5 = 13\n2x = 8\nx = 4"],
|
200 |
["Find the derivative of f(x)=xยฒ", "f'(x)=2x+1"],
|
201 |
+
["What is 15% of 200?", "0.15 ร 200 = 30"],
|
202 |
],
|
203 |
inputs=[q_in, s_in],
|
204 |
)
|