File size: 15,292 Bytes
f2532fa
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
import os
import traceback
import torch
from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel

DEFAULT_MODEL = "google/gemma-3-4b-it"
ADAPTER_PATH = "./gemma-lecture-adapter"
HUB_ADAPTER_ID = "noufwithy/gemma-lecture-adapter"

SUMMARIZE_SYSTEM_PROMPT = """You are a lecture summarization assistant.
Summarize the following lecture transcription into a comprehensive, structured summary with these sections:
- **Summary**: A concise overview of what the lecture covered
- **Key Points**: The main concepts, definitions, and important details covered in the lecture (use bullet points)
- **Action Points**: Any tasks, assignments, or follow-up actions mentioned by the lecturer

Cover ALL topics discussed. Do not omit any major points.
Output ONLY the summary. No explanations or extra commentary."""

# Quiz prompts match the training data format exactly (one question per call)
MCQ_SYSTEM_PROMPT = """You are an educational quiz generator.
Based on the following lecture transcription, generate a multiple choice question
with 4 options labeled A-D and indicate the correct answer.

Format:
Q1. [Question]
A) [Option]
B) [Option]
C) [Option]
D) [Option]
Correct Answer: [Letter]

Output ONLY the question. No explanations or extra commentary."""

SHORT_ANSWER_SYSTEM_PROMPT = """You are an educational quiz generator.
Based on the following lecture transcription, generate a short answer question
with the expected answer.

Format:
Q1. [Question]
Expected Answer: [Brief answer]

Output ONLY the question. No explanations or extra commentary."""

NUM_MCQ = 5
NUM_SHORT_ANSWER = 3

_model = None
_tokenizer = None


def _load_model(model_id: str = DEFAULT_MODEL, adapter_path: str = ADAPTER_PATH):
    global _model, _tokenizer
    if _model is not None:
        return _model, _tokenizer

    _tokenizer = AutoTokenizer.from_pretrained(model_id)

    # Try local adapter first, then HuggingFace Hub, then base model
    adapter_source = adapter_path if os.path.isdir(adapter_path) else HUB_ADAPTER_ID

    # Load in bfloat16 (bitsandbytes 4-bit/8-bit quantization broken with Gemma 3)
    try:
        print(f"Loading model with LoRA adapter from {adapter_source}...")
        base_model = AutoModelForCausalLM.from_pretrained(
            model_id,
            device_map="auto",
            dtype=torch.bfloat16,
            attn_implementation="eager",
        )
        _model = PeftModel.from_pretrained(base_model, adapter_source)
        _model.eval()
        print("LoRA adapter loaded successfully on bfloat16 base model.")
    except Exception as e:
        print(f"LoRA adapter failed ({e}), falling back to base model...")
        traceback.print_exc()
        _model = AutoModelForCausalLM.from_pretrained(
            model_id, device_map="auto", dtype=torch.bfloat16,
        )

    return _model, _tokenizer


def _generate(messages, max_new_tokens=2048, do_sample=False, temperature=0.7):
    """Generate text using model.generate() directly."""
    model, tokenizer = _load_model()

    # Format chat messages into a string, then tokenize
    prompt = tokenizer.apply_chat_template(
        messages, tokenize=False, add_generation_prompt=True
    )
    inputs = tokenizer(prompt, return_tensors="pt", add_special_tokens=False)
    input_ids = inputs["input_ids"].to(model.device)
    attention_mask = inputs["attention_mask"].to(model.device)

    print(f"[DEBUG] input length: {input_ids.shape[-1]} tokens")

    with torch.no_grad():
        outputs = model.generate(
            input_ids=input_ids,
            attention_mask=attention_mask,
            max_new_tokens=max_new_tokens,
            do_sample=do_sample,
            temperature=temperature if do_sample else None,
            top_p=0.9 if do_sample else None,
            repetition_penalty=1.3,
        )

    # Decode only the new tokens (skip the input)
    new_tokens = outputs[0][input_ids.shape[-1]:]
    print(f"[DEBUG] generated {len(new_tokens)} new tokens")

    response = tokenizer.decode(new_tokens, skip_special_tokens=True)
    return response.strip()


def _is_good_summary(text: str, transcript: str = "") -> bool:
    """Check if a summary meets minimum quality: long enough, not repetitive, not parroting."""
    if len(text) < 100:
        return False

    # Check for excessive repetition (same line or sentence repeated 2+ times)
    from collections import Counter
    for chunks in [
        [s.strip() for s in text.split("\n") if s.strip()],
        [s.strip() for s in text.split(".") if s.strip()],
    ]:
        if chunks:
            counts = Counter(chunks)
            most_common_count = counts.most_common(1)[0][1]
            if most_common_count >= 2:
                print(f"[QUALITY] Repetitive output detected ({most_common_count} repeats)")
                return False

    # Check if summary is just parroting the transcript (high word overlap)
    if transcript:
        summary_words = set(text.lower().split())
        transcript_words = set(transcript.lower().split())
        if summary_words and transcript_words:
            overlap = len(summary_words & transcript_words) / len(summary_words)
            if overlap > 0.85:
                print(f"[QUALITY] Summary too similar to transcript ({overlap:.0%} word overlap)")
                return False

    # Check if summary has enough key points (at least 3 bullet points)
    bullet_count = text.count("- ")
    has_key_points = "key points" in text.lower()
    if has_key_points and bullet_count < 3:
        print(f"[QUALITY] Summary has too few key points ({bullet_count})")
        return False

    # Check minimum unique content (summary should have substance)
    unique_lines = set(s.strip() for s in text.split("\n") if s.strip() and len(s.strip()) > 10)
    if len(unique_lines) < 5:
        print(f"[QUALITY] Summary too shallow ({len(unique_lines)} unique lines)")
        return False

    return True


def _generate_with_base_fallback(messages, transcript="", **kwargs):
    """Generate with adapter first. If output is bad, retry with base model."""
    result = _generate(messages, **kwargs)

    if _is_good_summary(result, transcript=transcript):
        return result

    # Adapter output is bad, try base model
    model, _ = _load_model()
    if isinstance(model, PeftModel):
        print("[FALLBACK] Adapter output too short or repetitive, retrying with base model...")
        model.disable_adapter_layers()
        try:
            result = _generate(messages, **kwargs)
        finally:
            model.enable_adapter_layers()
        print(f"[FALLBACK] base model response length: {len(result)}")

    return result


def _truncate_transcript(transcript: str, max_words: int = 4000) -> str:
    """Truncate transcript to fit model's effective context (trained on 3072 tokens)."""
    words = transcript.split()
    if len(words) <= max_words:
        return transcript
    print(f"[TRUNCATE] Transcript has {len(words)} words, truncating to {max_words}")
    return " ".join(words[:max_words])


def summarize_lecture(transcript: str, model: str = DEFAULT_MODEL) -> str:
    """Summarize a lecture transcript using Gemma."""
    if not transcript or not transcript.strip():
        return ""

    truncated = _truncate_transcript(transcript)
    messages = [
        {"role": "system", "content": SUMMARIZE_SYSTEM_PROMPT},
        {"role": "user", "content": f"Lecture transcription:\n\n{truncated}"},
    ]
    # Try adapter first, fall back to base model if quality is bad
    result = _generate_with_base_fallback(messages, transcript=transcript, do_sample=True, temperature=0.3)
    print(f"[DEBUG summarize] response length: {len(result)}")
    return result


def _extract_question_text(result: str) -> str:
    """Extract just the question text (first line after Q number) for dedup comparison."""
    import re
    match = re.search(r'Q\d+\.\s*(.+)', result)
    return match.group(1).strip().lower() if match else result.strip().lower()


def _is_good_quiz_answer(result: str, transcript: str = "") -> bool:
    """Check if a generated quiz question is reasonable quality."""
    # Reject if response doesn't match any expected format (no question generated)
    if "Correct Answer:" not in result and "Expected Answer:" not in result:
        print(f"[QUALITY] Response has no valid question format (missing Correct/Expected Answer)")
        return False

    # Reject if there's no actual question (Q1. pattern)
    if "Q1." not in result:
        print(f"[QUALITY] Response missing Q1. question marker")
        return False

    # Short answer: reject if expected answer is just a transcript fragment with no real content
    if "Expected Answer:" in result:
        answer = result.split("Expected Answer:")[-1].strip()
        # Reject vague/pointer answers like "right here", "this arrow", "at this point"
        vague_phrases = ["right here", "this arrow", "at this point", "this one", "over here", "right there"]
        if any(phrase in answer.lower() for phrase in vague_phrases):
            print(f"[QUALITY] Short answer too vague: {answer}")
            return False
        if len(answer.split()) < 2:
            print(f"[QUALITY] Short answer too short: {answer}")
            return False

    # MCQ: reject if it doesn't have 4 options or has duplicate options
    if "Correct Answer:" in result and "Expected Answer:" not in result:
        import re
        for label in ["A)", "B)", "C)", "D)"]:
            if label not in result:
                print(f"[QUALITY] MCQ missing option {label}")
                return False
        # Reject if options are mostly duplicated
        options = re.findall(r'[A-D]\)\s*(.+)', result)
        unique_options = set(opt.strip().lower() for opt in options)
        if len(unique_options) < 3:
            print(f"[QUALITY] MCQ has duplicate options ({len(unique_options)} unique out of {len(options)})")
            return False

    return True


def _dedup_mcq_options(result: str) -> str:
    """Remove duplicate MCQ options, keeping unique ones only."""
    import re
    options = re.findall(r'([A-D])\)\s*(.+)', result)
    if len(options) != 4:
        return result

    seen = {}
    unique = []
    for label, text in options:
        key = text.strip().lower()
        if key not in seen:
            seen[key] = True
            unique.append((label, text.strip()))

    if len(unique) == len(options):
        return result  # no duplicates

    print(f"[QUALITY] Removed {len(options) - len(unique)} duplicate MCQ option(s)")
    # Rebuild with correct labels
    lines = result.split("\n")
    new_lines = []
    option_idx = 0
    labels = ["A", "B", "C", "D"]
    for line in lines:
        if re.match(r'^[A-D]\)', line):
            if option_idx < len(unique):
                new_lines.append(f"{labels[option_idx]}) {unique[option_idx][1]}")
                option_idx += 1
        else:
            new_lines.append(line)

    return "\n".join(new_lines)


def _generate_quiz_with_fallback(messages, transcript="", **kwargs):
    """Generate a quiz question with adapter, fall back to base model if bad."""
    result = _generate(messages, **kwargs)

    if _is_good_quiz_answer(result, transcript):
        return result

    model, _ = _load_model()
    if isinstance(model, PeftModel):
        print("[FALLBACK] Quiz answer bad, retrying with base model...")
        model.disable_adapter_layers()
        try:
            result = _generate(messages, **kwargs)
        finally:
            model.enable_adapter_layers()

    return result


def _normalize_words(text: str) -> set[str]:
    """Strip punctuation from words for cleaner comparison."""
    import re
    return set(re.sub(r'[^\w\s]', '', word) for word in text.split() if word.strip())


def _is_duplicate(result: str, existing_parts: list[str]) -> bool:
    """Check if a generated question is too similar to any already generated."""
    new_q = _extract_question_text(result)
    for part in existing_parts:
        old_q = _extract_question_text(part)
        # Check if questions share most of their words (punctuation-stripped)
        new_words = _normalize_words(new_q)
        old_words = _normalize_words(old_q)
        if not new_words or not old_words:
            continue
        overlap = len(new_words & old_words) / min(len(new_words), len(old_words))
        if overlap > 0.7:
            print(f"[QUALITY] Duplicate question detected ({overlap:.0%} word overlap)")
            return True
    return False


def generate_quiz(transcript: str, model: str = DEFAULT_MODEL) -> str:
    """Generate quiz questions from a lecture transcript using Gemma.

    Generates questions one at a time to match training format, then combines them.
    Skips duplicate questions automatically.
    """
    if not transcript or not transcript.strip():
        return ""

    transcript = _truncate_transcript(transcript)
    parts = []
    max_retries = 2  # extra attempts per question if duplicate

    # Generate MCQs one at a time (matches training: one MCQ per example)
    for i in range(NUM_MCQ):
        print(f"[DEBUG quiz] generating MCQ {i + 1}/{NUM_MCQ}...")
        messages = [
            {"role": "system", "content": MCQ_SYSTEM_PROMPT},
            {"role": "user", "content": f"Lecture transcription:\n\n{transcript}"},
        ]
        good = False
        for attempt in range(1 + max_retries):
            result = _generate_quiz_with_fallback(messages, transcript=transcript, max_new_tokens=256, do_sample=True)
            if _is_good_quiz_answer(result, transcript) and not _is_duplicate(result, parts):
                good = True
                break
            print(f"[DEBUG quiz] MCQ {i + 1} attempt {attempt + 1} was bad or duplicate, retrying...")
        if good:
            result = _dedup_mcq_options(result)
            result = result.replace("Q1.", f"Q{len(parts) + 1}.", 1)
            parts.append(result)
        else:
            print(f"[DEBUG quiz] MCQ {i + 1} dropped (unreliable after {1 + max_retries} attempts)")

    # Generate short answer questions one at a time
    for i in range(NUM_SHORT_ANSWER):
        q_num = NUM_MCQ + i + 1
        print(f"[DEBUG quiz] generating short answer {i + 1}/{NUM_SHORT_ANSWER}...")
        messages = [
            {"role": "system", "content": SHORT_ANSWER_SYSTEM_PROMPT},
            {"role": "user", "content": f"Lecture transcription:\n\n{transcript}"},
        ]
        good = False
        for attempt in range(1 + max_retries):
            result = _generate_quiz_with_fallback(messages, transcript=transcript, max_new_tokens=256, do_sample=True)
            if _is_good_quiz_answer(result, transcript) and not _is_duplicate(result, parts):
                good = True
                break
            print(f"[DEBUG quiz] short answer {i + 1} attempt {attempt + 1} was bad or duplicate, retrying...")
        if good:
            result = result.replace("Q1.", f"Q{len(parts) + 1}.", 1)
            parts.append(result)
        else:
            print(f"[DEBUG quiz] short answer {i + 1} dropped (unreliable after {1 + max_retries} attempts)")

    combined = "\n\n".join(parts)
    print(f"[DEBUG quiz] total response length: {len(combined)}")
    return combined