File size: 14,341 Bytes
8205a2e |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 |
import gradio as gr
import torch
from transformers import AutoModelForCausalLM, AutoTokenizer, pipeline
from peft import PeftModel
import re
import json
from pathlib import Path
# ==================== CONFIGURATION ====================
# Base models
BASE_MODELS = {
"PHI-2 (2.7B)": "microsoft/phi-2",
"SmolLM2 (135M)": "HuggingFaceTB/SmolLM2-135M",
}
# Adapter configurations - update with your HuggingFace username
# Format: "username/repo-name" or local path
ADAPTERS = {
"PHI-2 (2.7B)": {
"No Fine-tuning (Base Model)": None,
"Baseline Fine-tuned": "CrystalRaindropsFall/phi2-gsm8k-baseline",
"Curriculum: Answer Length": "CrystalRaindropsFall/phi2-gsm8k-curriculum-answer-length",
"Curriculum: Complexity Score": "CrystalRaindropsFall/phi2-gsm8k-curriculum-complexity",
},
"SmolLM2 (135M)": {
"No Fine-tuning (Base Model)": None,
"Baseline Fine-tuned": "CrystalRaindropsFall/smolLM2-gsm8k-baseline",
"Curriculum: Answer Length": "CrystalRaindropsFall/smolLM2-gsm8k-curriculum-answer-length",
"Curriculum: Complexity Score": "CrystalRaindropsFall/smolLM2-gsm8k-curriculum-complexity",
},
}
# Sample math problems
SAMPLE_PROBLEMS = [
"Janet's ducks lay 16 eggs per day. She eats three for breakfast every morning and bakes muffins for her friends every day with four. She sells the remainder at the farmers' market daily for $2 per fresh duck egg. How much in dollars does she make every day at the farmers' market?",
"A robe takes 2 bolts of blue fiber and half that much white fiber. How many bolts in total does it take?",
"Josh decides to try flipping a house. He buys a house for $80,000 and then puts in $50,000 in repairs. This increased the value of the house by 150%. How much profit did he make?",
"James decides to run 3 sprints 3 times a week. He runs 60 meters each sprint. How many total meters does he run a week?",
"A store sells pencils for $0.50 each and notebooks for $3.00 each. If Sarah buys 6 pencils and 4 notebooks, how much does she spend in total?",
"Mike has 45 apples. He gives 1/3 of them to his friend and then buys 12 more apples. How many apples does Mike have now?",
"A train travels 120 miles in 2 hours. At the same speed, how far will it travel in 5 hours?",
]
# ==================== MODEL LOADING ====================
class ModelCache:
"""Cache loaded models to avoid reloading"""
def __init__(self):
self.current_base = None
self.current_adapter = None
self.model = None
self.tokenizer = None
self.pipe = None
def load_model(self, base_model_name, adapter_path=None):
"""Load model with optional adapter"""
cache_key = f"{base_model_name}_{adapter_path}"
current_key = f"{self.current_base}_{self.current_adapter}"
# Return cached if same
if cache_key == current_key and self.pipe is not None:
return self.pipe
# Clear old model
if self.model is not None:
del self.model
del self.tokenizer
del self.pipe
torch.cuda.empty_cache()
print(f"Loading {base_model_name}...")
# Load tokenizer
tokenizer = AutoTokenizer.from_pretrained(base_model_name)
if tokenizer.pad_token is None:
tokenizer.pad_token = tokenizer.eos_token
tokenizer.pad_token_id = tokenizer.eos_token_id
tokenizer.padding_side = "left"
# Load base model
model = AutoModelForCausalLM.from_pretrained(
base_model_name,
device_map="auto",
torch_dtype=torch.float16,
)
# Load adapter if specified
if adapter_path:
print(f"Loading adapter from {adapter_path}...")
if Path(adapter_path).exists():
# Local path
model = PeftModel.from_pretrained(model, adapter_path)
else:
# HuggingFace path
try:
model = PeftModel.from_pretrained(model, adapter_path)
except Exception as e:
print(f"Warning: Could not load adapter from {adapter_path}: {e}")
print("Using base model only")
# Create pipeline
pipe = pipeline(
"text-generation",
model=model,
tokenizer=tokenizer,
max_new_tokens=512,
do_sample=False, # Deterministic for math
pad_token_id=tokenizer.pad_token_id,
)
# Cache
self.current_base = base_model_name
self.current_adapter = adapter_path
self.model = model
self.tokenizer = tokenizer
self.pipe = pipe
return pipe
# Global cache
model_cache = ModelCache()
# ==================== HELPER FUNCTIONS ====================
def extract_answer(text):
"""Extract the final numerical answer from generated text"""
# Look for #### format (GSM8K style)
match = re.search(r"####\s*(-?\d+\.?\d*)", text)
if match:
return match.group(1).rstrip(".")
# Fallback: find last number
numbers = re.findall(r"-?\d+\.?\d*", text)
if numbers:
return numbers[-1].rstrip(".")
return "No answer found"
def format_solution(generated_text, question):
"""Format the solution for display"""
# Remove the question from the output (model echoes it)
solution = generated_text.replace(f"Question: {question}\nAnswer:", "").strip()
# Extract answer
final_answer = extract_answer(generated_text)
return solution, final_answer
# ==================== GRADIO INTERFACE ====================
def solve_math_problem(base_model, adapter_choice, question, max_tokens, temperature):
"""Main function to solve math problems"""
try:
# Get model path
base_model_path = BASE_MODELS[base_model]
adapter_path = ADAPTERS[base_model].get(adapter_choice)
# Load model
pipe = model_cache.load_model(base_model_path, adapter_path)
# Format prompt
prompt = f"Question: {question}\nAnswer:"
# Generate
outputs = pipe(
prompt,
max_new_tokens=max_tokens,
do_sample=temperature > 0,
temperature=temperature if temperature > 0 else None,
)
generated_text = outputs[0]["generated_text"]
# Format output
solution, final_answer = format_solution(generated_text, question)
# Create formatted output
output = f"""### Solution Steps:
{solution}
### Final Answer: **{final_answer}**
"""
return output
except Exception as e:
return f"โ Error: {str(e)}\n\nPlease check that the model and adapter are correctly loaded."
def update_adapter_choices(base_model):
"""Update adapter dropdown based on selected base model"""
adapters = list(ADAPTERS[base_model].keys())
return gr.Dropdown(choices=adapters, value=adapters[0])
def load_sample_problem(sample_idx):
"""Load a sample problem"""
if sample_idx is None or sample_idx >= len(SAMPLE_PROBLEMS):
return SAMPLE_PROBLEMS[0]
return SAMPLE_PROBLEMS[sample_idx]
# ==================== BUILD INTERFACE ====================
def create_demo():
"""Create the Gradio interface"""
with gr.Blocks(
theme=gr.themes.Soft(), title="Curriculum Design Matters: Math Reasoning Demo"
) as demo:
gr.Markdown(
"""
# ๐ Curriculum Design Matters: Training LLMs for Math Reasoning
<div style="font-size: 1.2em; line-height: 1.6;">
Compare how different training strategies affect mathematical reasoning in language models.
**Key Finding:** Not all curricula are equalโwrong curriculum design can hurt performance!
</div>
""",
elem_classes="header",
)
with gr.Row():
with gr.Column():
question_input = gr.Textbox(
lines=5,
placeholder="Enter a math word problem here...",
label="Enter Your Math Problem",
value=SAMPLE_PROBLEMS[0],
show_label=True,
)
with gr.Accordion("๐ Or Choose a Sample Problem", open=False):
sample_dropdown = gr.Dropdown(
choices=[
f"Sample {i + 1}: {prob[:50]}..."
for i, prob in enumerate(SAMPLE_PROBLEMS)
],
value=f"Sample 1: {SAMPLE_PROBLEMS[0][:50]}...",
label="Sample Problems",
scale=3,
)
load_sample_btn = gr.Button("๐ฅ Load Selected Sample", size="sm")
solve_btn = gr.Button("๐งฎ Solve Problem", variant="primary", size="lg")
gr.Markdown("### ๐ก Solution")
output_text = gr.Markdown(
value="*Solution will appear here after you click 'Solve Problem'...*",
label="Generated Solution",
)
gr.Markdown("### โ๏ธ Model Selection")
base_model = gr.Dropdown(
choices=list(BASE_MODELS.keys()),
value=list(BASE_MODELS.keys())[0],
label="Base Model",
info="Choose the foundation model",
)
adapter_choice = gr.Dropdown(
choices=list(ADAPTERS[list(BASE_MODELS.keys())[0]].keys()),
value=list(ADAPTERS[list(BASE_MODELS.keys())[0]].keys())[0],
label="Fine-tuning Strategy",
info="Choose training method",
)
with gr.Accordion("๐๏ธ Advanced Settings", open=False):
max_tokens = gr.Slider(
minimum=128,
maximum=512,
value=256,
step=32,
label="Max New Tokens",
info="Maximum length of solution",
)
temperature = gr.Slider(
minimum=0.0,
maximum=1.0,
value=0.0,
step=0.1,
label="Temperature",
info="0 = deterministic, >0 = creative",
)
# ==================== EVENT HANDLERS ====================
# Update adapters when base model changes
base_model.change(
fn=update_adapter_choices, inputs=[base_model], outputs=[adapter_choice]
)
# Load sample problem
def load_sample_fn(sample_name):
idx = int(sample_name.split()[1].split(":")[0]) - 1
return SAMPLE_PROBLEMS[idx]
load_sample_btn.click(
fn=load_sample_fn, inputs=[sample_dropdown], outputs=[question_input]
)
# Solve problem
solve_btn.click(
fn=solve_math_problem,
inputs=[
base_model,
adapter_choice,
question_input,
max_tokens,
temperature,
],
outputs=[output_text],
)
# ==================== BOTTOM INFO ====================
gr.Markdown("---")
with gr.Accordion("๐ Experimental Results & Key Findings", open=False):
gr.Markdown("""
### Results Summary
**PHI-2 (2.7B Parameters):**
- Baseline: 60.16% accuracy
- Curriculum (Answer Length): 59.38% (-0.78%) โ
- Curriculum (Complexity Score): 62.50% (+2.34%) โ
**SmolLM2 (135M Parameters):**
- Baseline: 2.15% accuracy
- Curriculum (Answer Length): 2.73% (+0.58%)
- Curriculum (Complexity Score): 2.93% (+0.78%)
### Key Insights
1. **Curriculum design is critical** - Wrong curriculum hurts performance
2. **Complexity matters more than length** - Steps ร operations beats simple answer length
3. **Model size affects benefits** - Larger models benefit more from curriculum learning
4. **Progressive difficulty works** - Easy โ Normal โ Difficult stages improve learning
""")
with gr.Accordion("๐ Training Methods Explained", open=False):
gr.Markdown("""
**No Fine-tuning:** Base model without any training on GSM8K
**Baseline Fine-tuned:** Standard fine-tuning on all problems at once
- All difficulty levels mixed together
- 3 epochs on full dataset
**Curriculum: Answer Length:** Progressive training based on solution length
- Stage 1 (Easy): Short solutions (< 100 chars)
- Stage 2 (Normal): Medium solutions (100-200 chars)
- Stage 3 (Difficult): Long solutions (> 200 chars)
- Result: Performance decreased! โ
**Curriculum: Complexity Score:** Progressive training based on steps ร operations
- Stage 1 (Easy): Few steps, simple operations
- Stage 2 (Normal): Moderate complexity
- Stage 3 (Difficult): Many steps, complex operations
- Result: Performance improved! โ
""")
with gr.Accordion("โน๏ธ About This Demo", open=False):
gr.Markdown("""
### Technical Details
**Models:**
- PHI-2: 2.7B parameter model by Microsoft
- SmolLM2: 135M parameter compact model by HuggingFace
**Dataset:** GSM8K (Grade School Math 8K) - 7,473 training and 1,319 test elementary school math word problems
**Training Method:** LoRA (Low-Rank Adaptation) fine-tuning
- Rank: 16, Alpha: 32
- Target modules: q_proj, k_proj, v_proj, o_proj
- 3 epochs per curriculum stage
- Learning rate: 3e-4
**Evaluation:** Exact match accuracy on GSM8K test set
### Links & Resources
๐ [GitHub Repository](#) | [Blog Post](#) | [Paper](#) | [Adapters on HuggingFace](#)
### Note
โ ๏ธ Models are loaded on-demand and cached in memory. First inference may take 30-60 seconds.
Models run on GPU if available, otherwise CPU (slower).
""")
return demo
# ==================== MAIN ====================
if __name__ == "__main__":
demo = create_demo()
demo.launch(
share=True, # Set to True to create public link
server_name="0.0.0.0", # Allow external access
server_port=7860,
show_error=True,
) |