import os
import re
import ast
import sys
import json
import time
from collections import deque
import gradio as gr
from transformers import pipeline
# ---------------------------------------------------------------------------
# Model Initialization — GPT-2 via HuggingFace Transformers
# ---------------------------------------------------------------------------
print("Loading GPT-2 model… this may take a moment on first run.")
generator = pipeline("text-generation", model="gpt2", max_new_tokens=400)
print("Model loaded.")
# ---------------------------------------------------------------------------
# MODULE 2 — Validation & Flow Management
# ---------------------------------------------------------------------------
MIN_DESCRIPTION_CHARS = 20
MAX_DESCRIPTION_CHARS = 3000
MIN_CODE_CHARS = 10
MAX_CODE_CHARS = 8000
MAX_CODE_LINES = 300
FORBIDDEN_PATTERNS = [
r"ignore (all |previous |above )?instructions",
r"disregard (all |previous |above )?instructions",
r"you are now",
r"act as (a |an )?",
r"<\s*(script|iframe|object|embed)",
r"system\s*prompt",
r"jailbreak",
]
PYTHON_KEYWORDS = {
"def", "class", "import", "from", "return", "if", "else", "elif",
"for", "while", "try", "except", "with", "lambda", "yield", "pass",
"raise", "assert", "in", "not", "and", "or", "True", "False", "None",
"print", "len", "range", "self",
}
class RateLimiter:
def __init__(self, max_calls: int = 5, window_seconds: int = 60):
self.max_calls = max_calls
self.window_seconds = window_seconds
self._timestamps: deque = deque()
def is_allowed(self) -> tuple[bool, str]:
now = time.time()
while self._timestamps and now - self._timestamps[0] > self.window_seconds:
self._timestamps.popleft()
if len(self._timestamps) >= self.max_calls:
wait = int(self.window_seconds - (now - self._timestamps[0])) + 1
return False, (
f"⏳ Rate limit reached — {self.max_calls} requests in "
f"{self.window_seconds}s. Please wait ~{wait}s and try again."
)
self._timestamps.append(now)
return True, ""
_rate_limiter = RateLimiter(max_calls=5, window_seconds=60)
def _check_forbidden(text: str) -> str | None:
lower = text.lower()
for pattern in FORBIDDEN_PATTERNS:
if re.search(pattern, lower):
return (
"Input contains disallowed content. "
"Please remove prompt-injection or HTML patterns and try again."
)
return None
def _looks_like_python(code: str) -> tuple[bool, str]:
tokens = set(re.findall(r"[A-Za-z_]\w*", code))
if not tokens.intersection(PYTHON_KEYWORDS):
return False, (
"🐍 The code doesn't appear to be Python — no recognisable Python "
"keywords found (e.g. def, class, import, return). "
"Please submit Python code only."
)
try:
ast.parse(code)
except SyntaxError as exc:
line_hint = f" (line {exc.lineno})" if exc.lineno else ""
return False, (
f"🐍 Python syntax error{line_hint}: {exc.msg}. "
"Please fix the syntax error before evaluating."
)
return True, ""
def validate_inputs(description: str, code: str) -> list[str]:
errors: list[str] = []
if not description or not description.strip():
errors.append("📋 Requirements description is required.")
if not code or not code.strip():
errors.append("🐍 Python code is required.")
if errors:
return errors
desc, code_ = description.strip(), code.strip()
if len(desc) < MIN_DESCRIPTION_CHARS:
errors.append(f"📋 Description too short ({len(desc)} chars) — minimum is {MIN_DESCRIPTION_CHARS} characters.")
if len(code_) < MIN_CODE_CHARS:
errors.append(f"🐍 Code too short ({len(code_)} chars) — minimum is {MIN_CODE_CHARS} characters.")
if len(desc) > MAX_DESCRIPTION_CHARS:
errors.append(f"📋 Description too long ({len(desc):,} chars) — max is {MAX_DESCRIPTION_CHARS:,} characters.")
if len(code_) > MAX_CODE_CHARS:
errors.append(f"🐍 Code too long ({len(code_):,} chars) — max is {MAX_CODE_CHARS:,} characters.")
if len(code_.splitlines()) > MAX_CODE_LINES:
errors.append(f"🐍 Code has too many lines ({len(code_.splitlines())}) — max is {MAX_CODE_LINES} lines.")
if err := _check_forbidden(desc):
errors.append(f"📋 {err}")
if err := _check_forbidden(code_):
errors.append(f"🐍 {err}")
if not errors:
is_python, py_error = _looks_like_python(code_)
if not is_python:
errors.append(py_error)
if not errors:
allowed, rate_msg = _rate_limiter.is_allowed()
if not allowed:
errors.append(rate_msg)
return errors
# ---------------------------------------------------------------------------
# MODULE 3 — Prompt Builder
# ---------------------------------------------------------------------------
def build_prompt(description: str, code: str) -> str:
return f"""You are a Python code reviewer. Evaluate whether the code meets the requirements.
Requirements:
{description}
Code:
```python
{code}
```
Reply ONLY with a valid JSON object (no extra text) using exactly these keys:
- "result": "Pass" or "Fail"
- "accuracy": integer 0-100
- "summary": one sentence explanation
- "issues": numbered list of issues as a single string, e.g. "1. Warning - missing docstring\\n2. Error - off-by-one"
JSON:
"""
# ---------------------------------------------------------------------------
# MODULE 4 — Generation & JSON Parsing
# ---------------------------------------------------------------------------
def _extract_json(text: str) -> dict:
"""Try to extract a JSON object from the raw model output."""
# Find the first { ... } block
match = re.search(r'\{.*?\}', text, re.DOTALL)
if match:
try:
return json.loads(match.group())
except json.JSONDecodeError:
pass
# Fallback: attempt to parse keys manually
result = "Pass" if re.search(r'"result"\s*:\s*"Pass"', text, re.I) else "Fail"
acc_m = re.search(r'"accuracy"\s*:\s*(\d+)', text)
accuracy = int(acc_m.group(1)) if acc_m else 50
sum_m = re.search(r'"summary"\s*:\s*"([^"]+)"', text)
summary = sum_m.group(1) if sum_m else "Unable to parse summary from model output."
iss_m = re.search(r'"issues"\s*:\s*"([^"]*)"', text, re.DOTALL)
issues = iss_m.group(1).replace("\\n", "\n") if iss_m else "No issues extracted."
return {"result": result, "accuracy": accuracy, "summary": summary, "issues": issues}
def generate_response(prompt: str) -> dict:
outputs = generator(prompt, do_sample=False, temperature=1.0)
raw_text = outputs[0]["generated_text"]
# Only look at the text appended after the prompt
new_text = raw_text[len(prompt):]
return _extract_json(new_text)
# ---------------------------------------------------------------------------
# Core orchestration
# ---------------------------------------------------------------------------
def format_for_display(raw: dict) -> tuple[str, str, str]:
emoji = "✅" if raw["result"].upper() == "PASS" else "❌"
verdict = f"{emoji} {raw['result']}"
acc_str = f"{raw['accuracy']}%" if raw.get("accuracy", -1) >= 0 else "N/A"
metrics = f"Accuracy: {acc_str}\n\nSummary: {raw['summary']}"
return verdict, metrics, raw.get("issues", "")
def validate_and_evaluate(description: str, code: str):
errors = validate_inputs(description, code)
if errors:
return "", "", "", "\n".join(f"{i+1}. {e}" for i, e in enumerate(errors))
prompt = build_prompt(description, code)
try:
raw = generate_response(prompt)
except Exception as exc:
return "", "", "", f"❌ Model error: {exc}"
verdict, metrics, issues = format_for_display(raw)
return verdict, metrics, issues, ""
# ---------------------------------------------------------------------------
# MODULE 1 — UI
# ---------------------------------------------------------------------------
CUSTOM_CSS = """
@import url('https://fonts.googleapis.com/css2?family=Space+Mono:wght@400;700&family=DM+Sans:ital,wght@0,300;0,400;0,500;0,600;1,400&display=swap');
*, *::before, *::after { box-sizing: border-box; }
body, .gradio-container {
font-family: 'DM Sans', sans-serif !important;
background: #0D0F14 !important;
color: #E8EAF0 !important;
}
.gradio-container {
max-width: 1140px !important;
margin: 0 auto !important;
padding: 32px 24px !important;
}
.eval-header {
display: flex; align-items: center; gap: 16px;
padding: 0 0 28px; border-bottom: 1px solid #2E3140; margin-bottom: 28px;
}
.eval-header .logo {
width: 44px; height: 44px; border-radius: 10px;
background: linear-gradient(135deg, #534AB7 0%, #1D9E75 100%);
display: flex; align-items: center; justify-content: center;
flex-shrink: 0; font-size: 22px; line-height: 1;
}
.eval-header h1 {
font-size: 20px !important; font-weight: 600 !important;
letter-spacing: -0.3px !important; color: #E8EAF0 !important; margin: 0 !important;
}
.eval-header p { font-size: 13px !important; color: #9DA0B0 !important; margin: 2px 0 0 !important; }
.eval-header .model-badge {
margin-left: auto; font-family: 'Space Mono', monospace; font-size: 10px;
background: #1E2028; border: 1px solid #3A3E52; color: #6A6E80;
padding: 4px 10px; border-radius: 4px; letter-spacing: 1.5px; white-space: nowrap;
}
.gradio-textbox textarea, .gradio-code textarea, .gradio-textbox input {
background: #161820 !important; border: 1px solid #2E3140 !important;
border-radius: 10px !important; color: #E8EAF0 !important;
font-family: 'DM Sans', sans-serif !important; font-size: 13.5px !important;
line-height: 1.7 !important; padding: 14px 16px !important;
transition: border-color 0.15s !important; resize: vertical !important;
}
.gradio-textbox textarea:focus, .gradio-code textarea:focus {
border-color: #534AB7 !important; outline: none !important;
box-shadow: 0 0 0 3px rgba(83, 74, 183, 0.15) !important;
}
.gradio-textbox textarea::placeholder { color: #4A4E60 !important; }
#code-input { min-height: 300px; }
#code-input .cm-editor { min-height: 300px; }
.gradio-textbox label span, .gradio-code label span {
font-family: 'DM Sans', sans-serif !important; font-size: 13px !important;
font-weight: 700 !important; color: #E8EAF0 !important;
text-transform: uppercase !important; letter-spacing: 0.8px !important;
}
#eval-btn {
background: #534AB7 !important; border: none !important; color: #fff !important;
font-family: 'DM Sans', sans-serif !important; font-size: 14px !important;
font-weight: 500 !important; padding: 12px 32px !important; border-radius: 8px !important;
cursor: pointer !important; transition: background 0.15s, transform 0.1s !important;
}
#eval-btn:hover { background: #7F77DD !important; transform: translateY(-1px) !important; }
#clear-btn {
background: transparent !important; border: 1px solid #3A3E52 !important;
color: #9DA0B0 !important; font-family: 'DM Sans', sans-serif !important;
font-size: 13px !important; padding: 12px 20px !important; border-radius: 8px !important;
cursor: pointer !important;
}
#clear-btn:hover { border-color: #7F77DD !important; color: #E8EAF0 !important; }
.results-heading {
font-size: 11px !important; font-weight: 500 !important; color: #6A6E80 !important;
text-transform: uppercase !important; letter-spacing: 1px !important;
padding: 0 0 16px !important; border-bottom: 1px solid #2E3140 !important; margin-bottom: 20px !important;
}
#verdict-out textarea, #accuracy-out textarea {
background: #161820 !important; border: 1px solid #2E3140 !important;
border-radius: 10px !important; font-family: 'Space Mono', monospace !important;
font-size: 26px !important; font-weight: 700 !important; text-align: center !important;
padding: 20px !important; color: #E8EAF0 !important; cursor: default !important;
}
#summary-out textarea, #issues-out textarea {
background: #161820 !important; border: 1px solid #2E3140 !important;
border-radius: 10px !important; font-family: 'DM Sans', sans-serif !important;
font-size: 13.5px !important; line-height: 1.7 !important; padding: 16px !important;
color: #C0C3D0 !important; cursor: default !important;
}
#issues-out textarea { min-height: 120px !important; line-height: 1.8 !important; }
#error-out textarea {
background: #1A0E0E !important; border: 1px solid #5a2020 !important;
border-radius: 10px !important; font-family: 'DM Sans', sans-serif !important;
font-size: 13.5px !important; padding: 14px 16px !important;
color: #F0997B !important; cursor: default !important;
}
.divider { height: 1px; background: #2E3140; margin: 20px 0; }
footer { display: none !important; }
::-webkit-scrollbar { width: 6px; height: 6px; }
::-webkit-scrollbar-track { background: transparent; }
::-webkit-scrollbar-thumb { background: #3A3E52; border-radius: 3px; }
::-webkit-scrollbar-thumb:hover { background: #534AB7; }
"""
HEADER_HTML = """
"""
RESULTS_HEADING_HTML = """
⬡ Evaluation Results
"""
INPUT_HINT_HTML = """
Tip — paste your description and code above, then click Evaluate.
Note: GPT-2 is a small model; results are best-effort and may need interpretation.
"""
def _split_metrics(metrics_str: str):
acc, summary = "", ""
if not metrics_str:
return acc, summary
for line in metrics_str.splitlines():
if line.startswith("Accuracy:"):
acc = line.replace("Accuracy:", "").strip()
elif line.startswith("Summary:"):
summary = line.replace("Summary:", "").strip()
return acc, summary
def _ui_evaluate(description: str, code: str):
verdict, metrics, issues, error = validate_and_evaluate(description, code)
accuracy, summary = _split_metrics(metrics)
if "PASS" in verdict.upper():
verdict_display = "✅ PASS"
elif "FAIL" in verdict.upper():
verdict_display = "❌ FAIL"
else:
verdict_display = verdict or ""
return verdict_display, accuracy, summary, issues, error
with gr.Blocks(title="Python Code Evaluator", css=CUSTOM_CSS) as app:
gr.HTML(HEADER_HTML)
with gr.Row(equal_height=True):
description_input = gr.Textbox(
label="Requirements Description",
placeholder=(
"Describe what the Python code is supposed to do…\n\n"
"Example: Write a function that accepts a list of integers "
"and returns the sum of all even numbers."
),
lines=10, max_lines=20, elem_id="desc-input",
)
code_input = gr.Code(
label="Python Code", language="python",
lines=10, max_lines=30, elem_id="code-input",
)
gr.HTML(INPUT_HINT_HTML)
with gr.Row():
eval_btn = gr.Button("Evaluate", variant="primary", elem_id="eval-btn", scale=0)
clear_btn = gr.Button("Clear", variant="secondary", elem_id="clear-btn", scale=0)
gr.HTML(RESULTS_HEADING_HTML)
with gr.Row(equal_height=True):
verdict_out = gr.Textbox(label="Verdict", interactive=False, elem_id="verdict-out", scale=1)
accuracy_out = gr.Textbox(label="Accuracy", interactive=False, elem_id="accuracy-out", scale=1)
summary_out = gr.Textbox(label="Summary", interactive=False, lines=3, elem_id="summary-out", scale=2)
issues_out = gr.Textbox(label="Issues Detected", interactive=False, lines=5, elem_id="issues-out")
error_out = gr.Textbox(label="Status / Errors", interactive=False, visible=True, elem_id="error-out")
outputs = [verdict_out, accuracy_out, summary_out, issues_out, error_out]
eval_btn.click(fn=_ui_evaluate, inputs=[description_input, code_input], outputs=outputs)
def _clear():
return "", "", "", "", "", ""
clear_btn.click(
fn=_clear, inputs=[],
outputs=[description_input, code_input, verdict_out, accuracy_out, summary_out, issues_out],
)
if __name__ == "__main__":
app.launch()