Update app.py
Browse files
app.py
CHANGED
|
@@ -1,7 +1,6 @@
|
|
| 1 |
"""
|
| 2 |
Qwen Coder Battle β Fine-tuned LoRA vs Base Model
|
| 3 |
===================================================
|
| 4 |
-
|
| 5 |
Compares Md-Asif/qwen-coder-python-ft (a LoRA adapter) against its base
|
| 6 |
Qwen2.5-Coder-3B model, side by side, on Python coding instructions.
|
| 7 |
|
|
@@ -13,6 +12,23 @@ Memory-efficient design for HF ZeroGPU free tier:
|
|
| 13 |
|
| 14 |
If you'd rather run this as two fully separate models, see the
|
| 15 |
`# ALTERNATE: two full models` comment block near the model loading code.
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 16 |
"""
|
| 17 |
|
| 18 |
import ast
|
|
@@ -28,21 +44,66 @@ from transformers import AutoModelForCausalLM, AutoTokenizer
|
|
| 28 |
# ----------------------------------------------------------------------------
|
| 29 |
# CONFIG β edit these to match how you actually trained the adapter
|
| 30 |
# ----------------------------------------------------------------------------
|
| 31 |
-
|
| 32 |
# β οΈ IMPORTANT: this MUST be the exact base model you fine-tuned on top of.
|
| 33 |
-
# If your adapter was trained on the
|
|
|
|
|
|
|
| 34 |
# "Qwen/Qwen2.5-Coder-3B"
|
|
|
|
|
|
|
| 35 |
BASE_MODEL_ID = "Qwen/Qwen2.5-Coder-3B"
|
| 36 |
-
|
| 37 |
ADAPTER_ID = "Md-Asif/qwen-coder-python-ft"
|
| 38 |
|
| 39 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 40 |
"You are an expert Python programmer. Given an instruction, respond with "
|
| 41 |
"clean, correct, well-documented Python code. Wrap code in a ```python "
|
| 42 |
"fenced block. Keep explanations brief."
|
| 43 |
)
|
| 44 |
|
| 45 |
MAX_NEW_TOKENS_DEFAULT = 512
|
|
|
|
| 46 |
DTYPE = torch.bfloat16
|
| 47 |
|
| 48 |
# ----------------------------------------------------------------------------
|
|
@@ -51,13 +112,15 @@ DTYPE = torch.bfloat16
|
|
| 51 |
# write as if a GPU is always present β the real GPU is attached only inside
|
| 52 |
# functions decorated with @spaces.GPU.
|
| 53 |
# ----------------------------------------------------------------------------
|
| 54 |
-
|
| 55 |
print("Loading tokenizer...")
|
| 56 |
try:
|
| 57 |
tokenizer = AutoTokenizer.from_pretrained(ADAPTER_ID)
|
| 58 |
except Exception:
|
| 59 |
tokenizer = AutoTokenizer.from_pretrained(BASE_MODEL_ID)
|
| 60 |
|
|
|
|
|
|
|
|
|
|
| 61 |
print("Loading base model (on CPU first β do NOT use device_map='auto' here;")
|
| 62 |
print("it forces real CUDA writes before ZeroGPU has attached a GPU)...")
|
| 63 |
base_model = AutoModelForCausalLM.from_pretrained(
|
|
@@ -91,11 +154,10 @@ model.eval()
|
|
| 91 |
# ----------------------------------------------------------------------------
|
| 92 |
# GENERATION
|
| 93 |
# ----------------------------------------------------------------------------
|
| 94 |
-
|
| 95 |
-
def build_inputs(instruction: str):
|
| 96 |
messages = [
|
| 97 |
-
{"role": "system", "content":
|
| 98 |
-
{"role": "user", "content":
|
| 99 |
]
|
| 100 |
text = tokenizer.apply_chat_template(
|
| 101 |
messages, tokenize=False, add_generation_prompt=True
|
|
@@ -103,13 +165,25 @@ def build_inputs(instruction: str):
|
|
| 103 |
return tokenizer(text, return_tensors="pt").to(model.device)
|
| 104 |
|
| 105 |
|
| 106 |
-
def
|
| 107 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 108 |
gen_kwargs = dict(
|
| 109 |
**inputs,
|
| 110 |
max_new_tokens=max_new_tokens,
|
| 111 |
do_sample=use_sampling,
|
| 112 |
-
pad_token_id=tokenizer.pad_token_id
|
| 113 |
)
|
| 114 |
if use_sampling:
|
| 115 |
gen_kwargs["temperature"] = max(temperature, 0.01)
|
|
@@ -122,25 +196,42 @@ def _generate(instruction: str, max_new_tokens: int, temperature: float, use_sam
|
|
| 122 |
|
| 123 |
new_tokens = output_ids[0][inputs["input_ids"].shape[1]:]
|
| 124 |
text = tokenizer.decode(new_tokens, skip_special_tokens=True)
|
| 125 |
-
|
| 126 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 127 |
|
| 128 |
|
| 129 |
@spaces.GPU(duration=90)
|
| 130 |
-
def run_battle(instruction: str, max_new_tokens: int, temperature: float, use_sampling: bool):
|
| 131 |
if not instruction or not instruction.strip():
|
| 132 |
raise gr.Error("Please enter an instruction first.")
|
| 133 |
|
| 134 |
-
|
| 135 |
-
|
|
|
|
|
|
|
|
|
|
| 136 |
|
| 137 |
-
# Base model (adapter disabled β same weights, no reload)
|
|
|
|
| 138 |
with model.disable_adapter():
|
| 139 |
-
base_text, base_time, base_tokens =
|
| 140 |
|
| 141 |
ft_metrics = analyze_response(ft_text, ft_time, ft_tokens)
|
| 142 |
base_metrics = analyze_response(base_text, base_time, base_tokens)
|
| 143 |
-
|
| 144 |
return ft_text, base_text, ft_metrics, base_metrics
|
| 145 |
|
| 146 |
|
|
@@ -151,7 +242,6 @@ def run_battle(instruction: str, max_new_tokens: int, temperature: float, use_sa
|
|
| 151 |
# against real test cases (e.g. HumanEval-style pass@1) β see the note in
|
| 152 |
# the README for how to extend this.
|
| 153 |
# ----------------------------------------------------------------------------
|
| 154 |
-
|
| 155 |
CODE_BLOCK_RE = re.compile(r"```(?:python)?\s*\n(.*?)```", re.DOTALL)
|
| 156 |
|
| 157 |
|
|
@@ -174,8 +264,8 @@ def analyze_response(text: str, elapsed: float, n_tokens: int) -> dict:
|
|
| 174 |
"num_classes": 0,
|
| 175 |
"has_docstring": False,
|
| 176 |
"lines_of_code": len([l for l in code.splitlines() if l.strip()]),
|
|
|
|
| 177 |
}
|
| 178 |
-
|
| 179 |
try:
|
| 180 |
tree = ast.parse(code)
|
| 181 |
metrics["syntax_valid"] = True
|
|
@@ -199,14 +289,12 @@ def analyze_response(text: str, elapsed: float, n_tokens: int) -> dict:
|
|
| 199 |
# Small speed bonus, capped, so a slow-but-correct answer isn't punished hard
|
| 200 |
score += min(10, metrics["tokens_per_sec"] / 5)
|
| 201 |
metrics["quality_score"] = round(min(score, 100), 1)
|
| 202 |
-
|
| 203 |
return metrics
|
| 204 |
|
| 205 |
|
| 206 |
# ----------------------------------------------------------------------------
|
| 207 |
# UI RENDERING HELPERS
|
| 208 |
# ----------------------------------------------------------------------------
|
| 209 |
-
|
| 210 |
def metric_row(label, ft_val, base_val, higher_is_better=True, suffix=""):
|
| 211 |
try:
|
| 212 |
ft_num, base_num = float(ft_val), float(base_val)
|
|
@@ -214,7 +302,6 @@ def metric_row(label, ft_val, base_val, higher_is_better=True, suffix=""):
|
|
| 214 |
base_win = base_num > ft_num if higher_is_better else base_num < ft_num
|
| 215 |
except (TypeError, ValueError):
|
| 216 |
ft_win = base_win = False
|
| 217 |
-
|
| 218 |
ft_cls = "metric-win" if ft_win else ""
|
| 219 |
base_cls = "metric-win" if base_win else ""
|
| 220 |
return f"""
|
|
@@ -228,7 +315,6 @@ def metric_row(label, ft_val, base_val, higher_is_better=True, suffix=""):
|
|
| 228 |
def render_comparison(ft_metrics: dict, base_metrics: dict) -> str:
|
| 229 |
winner = "Fine-tuned" if ft_metrics["quality_score"] >= base_metrics["quality_score"] else "Base"
|
| 230 |
winner_badge = f'<div class="winner-badge">π {winner} model scored higher</div>'
|
| 231 |
-
|
| 232 |
rows = "".join([
|
| 233 |
metric_row("Quality score", ft_metrics["quality_score"], base_metrics["quality_score"], suffix="/100"),
|
| 234 |
metric_row("Syntax valid", "β
Yes" if ft_metrics["syntax_valid"] else "β No",
|
|
@@ -242,7 +328,6 @@ def render_comparison(ft_metrics: dict, base_metrics: dict) -> str:
|
|
| 242 |
higher_is_better=False, suffix="s"),
|
| 243 |
metric_row("Tokens / sec", ft_metrics["tokens_per_sec"], base_metrics["tokens_per_sec"], suffix=" tok/s"),
|
| 244 |
])
|
| 245 |
-
|
| 246 |
return f"""
|
| 247 |
<div class="comparison-wrap">
|
| 248 |
{winner_badge}
|
|
@@ -254,7 +339,9 @@ def render_comparison(ft_metrics: dict, base_metrics: dict) -> str:
|
|
| 254 |
</table>
|
| 255 |
<div class="disclaimer">
|
| 256 |
Quality score is a static-analysis heuristic (syntax validity, structure,
|
| 257 |
-
docstrings, speed) β not a guarantee of runtime correctness.
|
|
|
|
|
|
|
| 258 |
</div>
|
| 259 |
</div>"""
|
| 260 |
|
|
@@ -262,14 +349,12 @@ def render_comparison(ft_metrics: dict, base_metrics: dict) -> str:
|
|
| 262 |
# ----------------------------------------------------------------------------
|
| 263 |
# CUSTOM STYLING
|
| 264 |
# ----------------------------------------------------------------------------
|
| 265 |
-
|
| 266 |
CUSTOM_CSS = """
|
| 267 |
:root {
|
| 268 |
--accent: #7c3aed;
|
| 269 |
--accent-2: #06b6d4;
|
| 270 |
}
|
| 271 |
.gradio-container { max-width: 1200px !important; margin: auto; }
|
| 272 |
-
|
| 273 |
#hero {
|
| 274 |
background: linear-gradient(135deg, #1e1b4b 0%, #4c1d95 45%, #0e7490 100%);
|
| 275 |
border-radius: 18px;
|
|
@@ -279,7 +364,6 @@ CUSTOM_CSS = """
|
|
| 279 |
}
|
| 280 |
#hero h1 { margin: 0 0 6px 0; font-size: 1.7rem; }
|
| 281 |
#hero p { margin: 0; opacity: 0.85; font-size: 0.95rem; }
|
| 282 |
-
|
| 283 |
.model-card {
|
| 284 |
border-radius: 14px;
|
| 285 |
padding: 4px;
|
|
@@ -297,7 +381,6 @@ CUSTOM_CSS = """
|
|
| 297 |
}
|
| 298 |
.label-ft { background: rgba(124,58,237,0.15); color: #a78bfa; }
|
| 299 |
.label-base { background: rgba(148,163,184,0.15); color: #94a3b8; }
|
| 300 |
-
|
| 301 |
.comparison-wrap { padding: 6px 4px; }
|
| 302 |
.winner-badge {
|
| 303 |
text-align: center;
|
|
@@ -318,27 +401,37 @@ CUSTOM_CSS = """
|
|
| 318 |
# ----------------------------------------------------------------------------
|
| 319 |
# GRADIO APP
|
| 320 |
# ----------------------------------------------------------------------------
|
| 321 |
-
|
| 322 |
EXAMPLES = [
|
| 323 |
-
"Write a function that checks if a string is a valid palindrome, ignoring punctuation and case.",
|
| 324 |
-
"Implement
|
| 325 |
-
"Write a script that reads a CSV file and returns the top 5 rows by a given column, using pandas.",
|
| 326 |
-
"Create a decorator that retries a function up to N times with exponential backoff.",
|
|
|
|
|
|
|
| 327 |
]
|
| 328 |
|
| 329 |
with gr.Blocks(css=CUSTOM_CSS, theme=gr.themes.Soft(primary_hue="violet", secondary_hue="cyan")) as demo:
|
| 330 |
gr.HTML("""
|
| 331 |
<div id="hero">
|
| 332 |
<h1>β‘ Qwen Coder Battle</h1>
|
| 333 |
-
<p>Md-Asif/qwen-coder-python-ft (LoRA fine-tune) vs Qwen2.5-Coder-3B (base) β
|
| 334 |
</div>
|
| 335 |
""")
|
| 336 |
|
| 337 |
-
|
| 338 |
-
|
| 339 |
-
|
| 340 |
-
|
| 341 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 342 |
|
| 343 |
with gr.Row():
|
| 344 |
run_btn = gr.Button("βοΈ Run the battle", variant="primary", scale=3)
|
|
@@ -348,7 +441,7 @@ with gr.Blocks(css=CUSTOM_CSS, theme=gr.themes.Soft(primary_hue="violet", second
|
|
| 348 |
use_sampling = gr.Checkbox(value=False, label="Use sampling (off = greedy, fairest for comparison)")
|
| 349 |
temperature = gr.Slider(0.1, 1.5, value=0.7, step=0.1, label="Temperature (only if sampling is on)")
|
| 350 |
|
| 351 |
-
gr.Examples(examples=EXAMPLES, inputs=instruction)
|
| 352 |
|
| 353 |
with gr.Row():
|
| 354 |
with gr.Column():
|
|
@@ -359,17 +452,16 @@ with gr.Blocks(css=CUSTOM_CSS, theme=gr.themes.Soft(primary_hue="violet", second
|
|
| 359 |
base_output = gr.Code(language="python", label=None, lines=18)
|
| 360 |
|
| 361 |
comparison_html = gr.HTML()
|
| 362 |
-
|
| 363 |
ft_metrics_state = gr.State()
|
| 364 |
base_metrics_state = gr.State()
|
| 365 |
|
| 366 |
-
def on_run(instr, max_tok, temp, sample):
|
| 367 |
-
ft_text, base_text, ft_m, base_m = run_battle(instr, max_tok, temp, sample)
|
| 368 |
return ft_text, base_text, render_comparison(ft_m, base_m), ft_m, base_m
|
| 369 |
|
| 370 |
run_btn.click(
|
| 371 |
on_run,
|
| 372 |
-
inputs=[instruction, max_tokens, temperature, use_sampling],
|
| 373 |
outputs=[ft_output, base_output, comparison_html, ft_metrics_state, base_metrics_state],
|
| 374 |
)
|
| 375 |
|
|
|
|
| 1 |
"""
|
| 2 |
Qwen Coder Battle β Fine-tuned LoRA vs Base Model
|
| 3 |
===================================================
|
|
|
|
| 4 |
Compares Md-Asif/qwen-coder-python-ft (a LoRA adapter) against its base
|
| 5 |
Qwen2.5-Coder-3B model, side by side, on Python coding instructions.
|
| 6 |
|
|
|
|
| 12 |
|
| 13 |
If you'd rather run this as two fully separate models, see the
|
| 14 |
`# ALTERNATE: two full models` comment block near the model loading code.
|
| 15 |
+
|
| 16 |
+
--------------------------------------------------------------------------
|
| 17 |
+
IMPORTANT β prompt format
|
| 18 |
+
--------------------------------------------------------------------------
|
| 19 |
+
Md-Asif/python-fine-tune trains on FOUR distinct (system_prompt, task_tag)
|
| 20 |
+
pairs, keyed by a `task` column: GENERATE, REFACTOR, DEBUG, CODE_REVIEW.
|
| 21 |
+
Every training example's user turn is prefixed with "<TASK:XXXX>\n\n".
|
| 22 |
+
|
| 23 |
+
The LoRA adapter has only ever seen instructions in that exact shape. Feed
|
| 24 |
+
it a plain instruction with a different system prompt and no tag, and it's
|
| 25 |
+
operating out-of-distribution β often producing WORSE output than the
|
| 26 |
+
untouched base model, which was never trained to expect the tag at all.
|
| 27 |
+
|
| 28 |
+
So: the fine-tuned model gets the task-specific system prompt + tagged
|
| 29 |
+
instruction. The base model gets a plain, generic system prompt and the
|
| 30 |
+
RAW instruction (no tag) β that's the fair, apples-to-apples comparison,
|
| 31 |
+
since neither format is "natural" to the base model anyway.
|
| 32 |
"""
|
| 33 |
|
| 34 |
import ast
|
|
|
|
| 44 |
# ----------------------------------------------------------------------------
|
| 45 |
# CONFIG β edit these to match how you actually trained the adapter
|
| 46 |
# ----------------------------------------------------------------------------
|
|
|
|
| 47 |
# β οΈ IMPORTANT: this MUST be the exact base model you fine-tuned on top of.
|
| 48 |
+
# If your adapter was trained on the Instruct checkpoint, use:
|
| 49 |
+
# "Qwen/Qwen2.5-Coder-3B-Instruct"
|
| 50 |
+
# If it was trained on the plain base, use:
|
| 51 |
# "Qwen/Qwen2.5-Coder-3B"
|
| 52 |
+
# Getting this wrong silently wrecks fine-tuned quality β the LoRA delta
|
| 53 |
+
# was computed relative to ONE specific set of base weights.
|
| 54 |
BASE_MODEL_ID = "Qwen/Qwen2.5-Coder-3B"
|
|
|
|
| 55 |
ADAPTER_ID = "Md-Asif/qwen-coder-python-ft"
|
| 56 |
|
| 57 |
+
# System prompts, EXACTLY matching what's in the training dataset's
|
| 58 |
+
# `messages[0]["content"]` for each task. Keep these byte-identical to the
|
| 59 |
+
# dataset β even small wording changes are a mild distribution shift.
|
| 60 |
+
TASK_SYSTEM_PROMPTS = {
|
| 61 |
+
"GENERATE": (
|
| 62 |
+
"You are a senior Python engineer.\n"
|
| 63 |
+
"Write complete, correct, production-ready Python code.\n"
|
| 64 |
+
"Include type hints, docstrings, and handle edge cases.\n"
|
| 65 |
+
"No placeholders. No TODOs. Only working code."
|
| 66 |
+
),
|
| 67 |
+
"REFACTOR": (
|
| 68 |
+
"You are an expert Python refactoring assistant.\n"
|
| 69 |
+
"Rewrite the code following:\n"
|
| 70 |
+
"- PEP 8\n"
|
| 71 |
+
"- Type hints\n"
|
| 72 |
+
"- Docstrings\n"
|
| 73 |
+
"- Best practices\n"
|
| 74 |
+
"- SOLID principles\n\n"
|
| 75 |
+
"Do not change behavior. Only improve structure and quality."
|
| 76 |
+
),
|
| 77 |
+
"DEBUG": (
|
| 78 |
+
"You are an expert Python debugging engineer.\n"
|
| 79 |
+
"Locate the exact bug.\n"
|
| 80 |
+
"Explain precisely why it happens.\n"
|
| 81 |
+
"Produce corrected, complete Python code."
|
| 82 |
+
),
|
| 83 |
+
"CODE_REVIEW": (
|
| 84 |
+
"You are a senior Python code reviewer.\n"
|
| 85 |
+
"Review code like a staff engineer.\n\n"
|
| 86 |
+
"Find and fix:\n"
|
| 87 |
+
"- Syntax bugs\n"
|
| 88 |
+
"- Runtime bugs\n"
|
| 89 |
+
"- Logic bugs\n"
|
| 90 |
+
"- Performance issues\n"
|
| 91 |
+
"- Security issues\n"
|
| 92 |
+
"- Maintainability issues\n"
|
| 93 |
+
"Produce corrected production-quality Python code and explain every change."
|
| 94 |
+
),
|
| 95 |
+
}
|
| 96 |
+
|
| 97 |
+
# What the BASE model gets instead β a neutral prompt with no dataset-specific
|
| 98 |
+
# tagging, since the base model was never trained on the <TASK:X> scheme.
|
| 99 |
+
BASE_SYSTEM_PROMPT = (
|
| 100 |
"You are an expert Python programmer. Given an instruction, respond with "
|
| 101 |
"clean, correct, well-documented Python code. Wrap code in a ```python "
|
| 102 |
"fenced block. Keep explanations brief."
|
| 103 |
)
|
| 104 |
|
| 105 |
MAX_NEW_TOKENS_DEFAULT = 512
|
| 106 |
+
WARMUP_TOKENS = 8 # tiny generation to absorb CUDA cold-start before timing
|
| 107 |
DTYPE = torch.bfloat16
|
| 108 |
|
| 109 |
# ----------------------------------------------------------------------------
|
|
|
|
| 112 |
# write as if a GPU is always present β the real GPU is attached only inside
|
| 113 |
# functions decorated with @spaces.GPU.
|
| 114 |
# ----------------------------------------------------------------------------
|
|
|
|
| 115 |
print("Loading tokenizer...")
|
| 116 |
try:
|
| 117 |
tokenizer = AutoTokenizer.from_pretrained(ADAPTER_ID)
|
| 118 |
except Exception:
|
| 119 |
tokenizer = AutoTokenizer.from_pretrained(BASE_MODEL_ID)
|
| 120 |
|
| 121 |
+
if tokenizer.pad_token_id is None:
|
| 122 |
+
tokenizer.pad_token = tokenizer.eos_token
|
| 123 |
+
|
| 124 |
print("Loading base model (on CPU first β do NOT use device_map='auto' here;")
|
| 125 |
print("it forces real CUDA writes before ZeroGPU has attached a GPU)...")
|
| 126 |
base_model = AutoModelForCausalLM.from_pretrained(
|
|
|
|
| 154 |
# ----------------------------------------------------------------------------
|
| 155 |
# GENERATION
|
| 156 |
# ----------------------------------------------------------------------------
|
| 157 |
+
def _build_inputs(system_prompt: str, user_content: str):
|
|
|
|
| 158 |
messages = [
|
| 159 |
+
{"role": "system", "content": system_prompt},
|
| 160 |
+
{"role": "user", "content": user_content},
|
| 161 |
]
|
| 162 |
text = tokenizer.apply_chat_template(
|
| 163 |
messages, tokenize=False, add_generation_prompt=True
|
|
|
|
| 165 |
return tokenizer(text, return_tensors="pt").to(model.device)
|
| 166 |
|
| 167 |
|
| 168 |
+
def build_inputs_ft(instruction: str, task: str):
|
| 169 |
+
"""Fine-tuned path: task-specific system prompt + <TASK:X> tagged instruction,
|
| 170 |
+
matching the training data format exactly."""
|
| 171 |
+
system_prompt = TASK_SYSTEM_PROMPTS[task]
|
| 172 |
+
tagged_instruction = f"<TASK:{task}>\n\n{instruction}"
|
| 173 |
+
return _build_inputs(system_prompt, tagged_instruction)
|
| 174 |
+
|
| 175 |
+
|
| 176 |
+
def build_inputs_base(instruction: str):
|
| 177 |
+
"""Base path: neutral system prompt, raw instruction, no tag."""
|
| 178 |
+
return _build_inputs(BASE_SYSTEM_PROMPT, instruction)
|
| 179 |
+
|
| 180 |
+
|
| 181 |
+
def _run_generate(inputs, max_new_tokens: int, temperature: float, use_sampling: bool):
|
| 182 |
gen_kwargs = dict(
|
| 183 |
**inputs,
|
| 184 |
max_new_tokens=max_new_tokens,
|
| 185 |
do_sample=use_sampling,
|
| 186 |
+
pad_token_id=tokenizer.pad_token_id,
|
| 187 |
)
|
| 188 |
if use_sampling:
|
| 189 |
gen_kwargs["temperature"] = max(temperature, 0.01)
|
|
|
|
| 196 |
|
| 197 |
new_tokens = output_ids[0][inputs["input_ids"].shape[1]:]
|
| 198 |
text = tokenizer.decode(new_tokens, skip_special_tokens=True)
|
| 199 |
+
return text, elapsed, len(new_tokens)
|
| 200 |
+
|
| 201 |
+
|
| 202 |
+
def _warmup():
|
| 203 |
+
"""Absorb CUDA context attach / kernel autotune cost with a throwaway
|
| 204 |
+
generation, so neither the FT nor the base timing eats the cold-start
|
| 205 |
+
penalty. Must run once per @spaces.GPU-decorated call, before any timed
|
| 206 |
+
generation."""
|
| 207 |
+
dummy = _build_inputs(BASE_SYSTEM_PROMPT, "print hello world")
|
| 208 |
+
with torch.no_grad():
|
| 209 |
+
model.generate(
|
| 210 |
+
**dummy,
|
| 211 |
+
max_new_tokens=WARMUP_TOKENS,
|
| 212 |
+
do_sample=False,
|
| 213 |
+
pad_token_id=tokenizer.pad_token_id,
|
| 214 |
+
)
|
| 215 |
|
| 216 |
|
| 217 |
@spaces.GPU(duration=90)
|
| 218 |
+
def run_battle(instruction: str, task: str, max_new_tokens: int, temperature: float, use_sampling: bool):
|
| 219 |
if not instruction or not instruction.strip():
|
| 220 |
raise gr.Error("Please enter an instruction first.")
|
| 221 |
|
| 222 |
+
_warmup()
|
| 223 |
+
|
| 224 |
+
# Fine-tuned model (adapter active, task-specific prompt + tag)
|
| 225 |
+
ft_inputs = build_inputs_ft(instruction, task)
|
| 226 |
+
ft_text, ft_time, ft_tokens = _run_generate(ft_inputs, max_new_tokens, temperature, use_sampling)
|
| 227 |
|
| 228 |
+
# Base model (adapter disabled β same weights, no reload β neutral prompt)
|
| 229 |
+
base_inputs = build_inputs_base(instruction)
|
| 230 |
with model.disable_adapter():
|
| 231 |
+
base_text, base_time, base_tokens = _run_generate(base_inputs, max_new_tokens, temperature, use_sampling)
|
| 232 |
|
| 233 |
ft_metrics = analyze_response(ft_text, ft_time, ft_tokens)
|
| 234 |
base_metrics = analyze_response(base_text, base_time, base_tokens)
|
|
|
|
| 235 |
return ft_text, base_text, ft_metrics, base_metrics
|
| 236 |
|
| 237 |
|
|
|
|
| 242 |
# against real test cases (e.g. HumanEval-style pass@1) β see the note in
|
| 243 |
# the README for how to extend this.
|
| 244 |
# ----------------------------------------------------------------------------
|
|
|
|
| 245 |
CODE_BLOCK_RE = re.compile(r"```(?:python)?\s*\n(.*?)```", re.DOTALL)
|
| 246 |
|
| 247 |
|
|
|
|
| 264 |
"num_classes": 0,
|
| 265 |
"has_docstring": False,
|
| 266 |
"lines_of_code": len([l for l in code.splitlines() if l.strip()]),
|
| 267 |
+
"raw_text": text,
|
| 268 |
}
|
|
|
|
| 269 |
try:
|
| 270 |
tree = ast.parse(code)
|
| 271 |
metrics["syntax_valid"] = True
|
|
|
|
| 289 |
# Small speed bonus, capped, so a slow-but-correct answer isn't punished hard
|
| 290 |
score += min(10, metrics["tokens_per_sec"] / 5)
|
| 291 |
metrics["quality_score"] = round(min(score, 100), 1)
|
|
|
|
| 292 |
return metrics
|
| 293 |
|
| 294 |
|
| 295 |
# ----------------------------------------------------------------------------
|
| 296 |
# UI RENDERING HELPERS
|
| 297 |
# ----------------------------------------------------------------------------
|
|
|
|
| 298 |
def metric_row(label, ft_val, base_val, higher_is_better=True, suffix=""):
|
| 299 |
try:
|
| 300 |
ft_num, base_num = float(ft_val), float(base_val)
|
|
|
|
| 302 |
base_win = base_num > ft_num if higher_is_better else base_num < ft_num
|
| 303 |
except (TypeError, ValueError):
|
| 304 |
ft_win = base_win = False
|
|
|
|
| 305 |
ft_cls = "metric-win" if ft_win else ""
|
| 306 |
base_cls = "metric-win" if base_win else ""
|
| 307 |
return f"""
|
|
|
|
| 315 |
def render_comparison(ft_metrics: dict, base_metrics: dict) -> str:
|
| 316 |
winner = "Fine-tuned" if ft_metrics["quality_score"] >= base_metrics["quality_score"] else "Base"
|
| 317 |
winner_badge = f'<div class="winner-badge">π {winner} model scored higher</div>'
|
|
|
|
| 318 |
rows = "".join([
|
| 319 |
metric_row("Quality score", ft_metrics["quality_score"], base_metrics["quality_score"], suffix="/100"),
|
| 320 |
metric_row("Syntax valid", "β
Yes" if ft_metrics["syntax_valid"] else "β No",
|
|
|
|
| 328 |
higher_is_better=False, suffix="s"),
|
| 329 |
metric_row("Tokens / sec", ft_metrics["tokens_per_sec"], base_metrics["tokens_per_sec"], suffix=" tok/s"),
|
| 330 |
])
|
|
|
|
| 331 |
return f"""
|
| 332 |
<div class="comparison-wrap">
|
| 333 |
{winner_badge}
|
|
|
|
| 339 |
</table>
|
| 340 |
<div class="disclaimer">
|
| 341 |
Quality score is a static-analysis heuristic (syntax validity, structure,
|
| 342 |
+
docstrings, speed) β not a guarantee of runtime correctness. Fine-tuned
|
| 343 |
+
gets its trained system prompt + <TASK:{{task}}> tag; base gets a
|
| 344 |
+
neutral prompt with no tag β matching how each was actually trained.
|
| 345 |
</div>
|
| 346 |
</div>"""
|
| 347 |
|
|
|
|
| 349 |
# ----------------------------------------------------------------------------
|
| 350 |
# CUSTOM STYLING
|
| 351 |
# ----------------------------------------------------------------------------
|
|
|
|
| 352 |
CUSTOM_CSS = """
|
| 353 |
:root {
|
| 354 |
--accent: #7c3aed;
|
| 355 |
--accent-2: #06b6d4;
|
| 356 |
}
|
| 357 |
.gradio-container { max-width: 1200px !important; margin: auto; }
|
|
|
|
| 358 |
#hero {
|
| 359 |
background: linear-gradient(135deg, #1e1b4b 0%, #4c1d95 45%, #0e7490 100%);
|
| 360 |
border-radius: 18px;
|
|
|
|
| 364 |
}
|
| 365 |
#hero h1 { margin: 0 0 6px 0; font-size: 1.7rem; }
|
| 366 |
#hero p { margin: 0; opacity: 0.85; font-size: 0.95rem; }
|
|
|
|
| 367 |
.model-card {
|
| 368 |
border-radius: 14px;
|
| 369 |
padding: 4px;
|
|
|
|
| 381 |
}
|
| 382 |
.label-ft { background: rgba(124,58,237,0.15); color: #a78bfa; }
|
| 383 |
.label-base { background: rgba(148,163,184,0.15); color: #94a3b8; }
|
|
|
|
| 384 |
.comparison-wrap { padding: 6px 4px; }
|
| 385 |
.winner-badge {
|
| 386 |
text-align: center;
|
|
|
|
| 401 |
# ----------------------------------------------------------------------------
|
| 402 |
# GRADIO APP
|
| 403 |
# ----------------------------------------------------------------------------
|
|
|
|
| 404 |
EXAMPLES = [
|
| 405 |
+
["GENERATE", "Write a function that checks if a string is a valid palindrome, ignoring punctuation and case."],
|
| 406 |
+
["GENERATE", "Implement an LRU cache from scratch without using functools.lru_cache."],
|
| 407 |
+
["GENERATE", "Write a script that reads a CSV file and returns the top 5 rows by a given column, using pandas."],
|
| 408 |
+
["GENERATE", "Create a decorator that retries a function up to N times with exponential backoff."],
|
| 409 |
+
["DEBUG", "def add(a, b):\n return a + b\n\nprint(add(2))\n\n# This raises a TypeError. Find and fix the bug."],
|
| 410 |
+
["REFACTOR", "def f(x):\n y=[]\n for i in range(len(x)):\n if x[i]%2==0:\n y.append(x[i])\n return y"],
|
| 411 |
]
|
| 412 |
|
| 413 |
with gr.Blocks(css=CUSTOM_CSS, theme=gr.themes.Soft(primary_hue="violet", secondary_hue="cyan")) as demo:
|
| 414 |
gr.HTML("""
|
| 415 |
<div id="hero">
|
| 416 |
<h1>β‘ Qwen Coder Battle</h1>
|
| 417 |
+
<p>Md-Asif/qwen-coder-python-ft (LoRA fine-tune) vs Qwen2.5-Coder-3B (base) β each prompted the way it was actually trained. Judge for yourself.</p>
|
| 418 |
</div>
|
| 419 |
""")
|
| 420 |
|
| 421 |
+
with gr.Row():
|
| 422 |
+
instruction = gr.Textbox(
|
| 423 |
+
label="Coding instruction",
|
| 424 |
+
placeholder="e.g. Write a function that merges two sorted lists in O(n) time.",
|
| 425 |
+
lines=3,
|
| 426 |
+
scale=3,
|
| 427 |
+
)
|
| 428 |
+
task = gr.Dropdown(
|
| 429 |
+
choices=["GENERATE", "REFACTOR", "DEBUG", "CODE_REVIEW"],
|
| 430 |
+
value="GENERATE",
|
| 431 |
+
label="Task type",
|
| 432 |
+
info="Must match the dataset's <TASK:X> categories for a fair fine-tuned comparison.",
|
| 433 |
+
scale=1,
|
| 434 |
+
)
|
| 435 |
|
| 436 |
with gr.Row():
|
| 437 |
run_btn = gr.Button("βοΈ Run the battle", variant="primary", scale=3)
|
|
|
|
| 441 |
use_sampling = gr.Checkbox(value=False, label="Use sampling (off = greedy, fairest for comparison)")
|
| 442 |
temperature = gr.Slider(0.1, 1.5, value=0.7, step=0.1, label="Temperature (only if sampling is on)")
|
| 443 |
|
| 444 |
+
gr.Examples(examples=EXAMPLES, inputs=[task, instruction])
|
| 445 |
|
| 446 |
with gr.Row():
|
| 447 |
with gr.Column():
|
|
|
|
| 452 |
base_output = gr.Code(language="python", label=None, lines=18)
|
| 453 |
|
| 454 |
comparison_html = gr.HTML()
|
|
|
|
| 455 |
ft_metrics_state = gr.State()
|
| 456 |
base_metrics_state = gr.State()
|
| 457 |
|
| 458 |
+
def on_run(instr, task_choice, max_tok, temp, sample):
|
| 459 |
+
ft_text, base_text, ft_m, base_m = run_battle(instr, task_choice, max_tok, temp, sample)
|
| 460 |
return ft_text, base_text, render_comparison(ft_m, base_m), ft_m, base_m
|
| 461 |
|
| 462 |
run_btn.click(
|
| 463 |
on_run,
|
| 464 |
+
inputs=[instruction, task, max_tokens, temperature, use_sampling],
|
| 465 |
outputs=[ft_output, base_output, comparison_html, ft_metrics_state, base_metrics_state],
|
| 466 |
)
|
| 467 |
|