File size: 18,550 Bytes
db06e07
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
# app.py
# Single-page Gradio app for Hugging Face Spaces
# - Trains MiniGPT and classifier on startup (tiny datasets, short epochs by default)
# - Large, centered UI with three panels:
#     1) Instruction -> Response
#     2) Sentiment Classification
#     3) Next word + dataset sentence completion (prefix of two words)
# - Instant input moderation: banned words trigger immediate error and block
# - Greedy decoding for stable minimal outputs

import math, re, os, torch, torch.nn as nn
from torch.utils.data import Dataset, DataLoader
import gradio as gr

# ----------------------------
# 1) Data preparation
# ----------------------------
lm_corpus = [
    "the cat sits on the mat",
    "the dog chases the ball",
    "a small model can learn patterns",
    "language models predict next tokens",
    "transformers use attention mechanism",
    "training on tiny data is limited",
    "we build a model from scratch",
    "this is a minimal example",
    "positional embeddings encode order",
    "causal masking prevents peeking ahead",
]

cls_data = [
    ("this is bad", 0),
    ("i dislike this", 0),
    ("terrible and awful", 0),
    ("this is good", 1),
    ("i like this", 1),
    ("wonderful and great", 1),
]

inst_data_base = [
    ("<INSTR> write a short greeting <ENDINSTR>", "<RESP> hello! <ENDRESP>"),
    ("<INSTR> answer briefly what is a cat <ENDINSTR>", "<RESP> a small animal. <ENDRESP>"),
    ("<INSTR> continue the sun is <ENDINSTR>", "<RESP> bright. <ENDRESP>"),
]
inst_data = inst_data_base * 64  # stabilize tiny-data learning

# ----------------------------
# Tokenization (word-level)
# ----------------------------
def normalize_text(s):
    s = s.lower().strip()
    s = re.sub(r'([.!?,:;])', r' \1 ', s)
    s = re.sub(r'\s+', ' ', s)
    return s

def build_vocab(texts):
    tokens = set()
    specials = ["<pad>", "<bos>", "<eos>"]
    for t in texts:
        t = normalize_text(t)
        for tok in t.split():
            tokens.add(tok)
    vocab = specials + sorted(list(tokens))
    stoi = {s: i for i, s in enumerate(vocab)}
    itos = {i: s for s, i in stoi.items()}
    return vocab, stoi, itos

all_texts = lm_corpus + [x for x,_ in cls_data] + [a for a,_ in inst_data_base] + [b for _,b in inst_data_base]
vocab, stoi, itos = build_vocab(all_texts)
PAD, BOS, EOS = stoi["<pad>"], stoi["<bos>"], stoi["<eos>"]
vocab_size = len(vocab)

def encode(text, max_len=None, add_special=True):
    text = normalize_text(text)
    toks = text.split()
    ids = ([BOS] if add_special else []) + [stoi.get(tok, PAD) for tok in toks] + ([EOS] if add_special else [])
    if max_len is not None:
        ids = ids[:max_len]
        if len(ids) < max_len:
            ids = ids + [PAD] * (max_len - len(ids))
    return torch.tensor(ids, dtype=torch.long)

def decode(ids):
    toks = [itos.get(i, "") for i in ids]
    toks = [t for t in toks if t not in ("<pad>", "<bos>", "<eos>")]
    out = " ".join(toks)
    out = re.sub(r'\s+([.!?,:;])', r'\1', out)
    return out.strip()

# ----------------------------
# Datasets
# ----------------------------
class LMPretrainDataset(Dataset):
    def __init__(self, texts, block_size=64):
        self.samples = []
        for t in texts:
            ids = encode(t, max_len=block_size, add_special=True)
            self.samples.append((ids[:-1], ids[1:]))
    def __len__(self): return len(self.samples)
    def __getitem__(self, idx): return self.samples[idx]

class ClassificationDataset(Dataset):
    def __init__(self, pairs, block_size=64):
        self.samples = []
        for text, label in pairs:
            ids = encode(text, max_len=block_size, add_special=True)
            self.samples.append((ids, torch.tensor(label, dtype=torch.long)))
    def __len__(self): return len(self.samples)
    def __getitem__(self, idx): return self.samples[idx]

class InstructionDataset(Dataset):
    def __init__(self, pairs, block_size=64):
        self.samples = []
        for instr, resp in pairs:
            instr_ids = encode(instr, add_special=False).tolist()
            resp_ids  = encode(resp, add_special=False).tolist()
            seq = [BOS] + instr_ids + [EOS] + [BOS] + resp_ids + [EOS]
            seq = seq[:block_size]
            if len(seq) < block_size: seq += [PAD] * (block_size - len(seq))
            ids = torch.tensor(seq, dtype=torch.long)
            self.samples.append((ids[:-1], ids[1:]))
    def __len__(self): return len(self.samples)
    def __getitem__(self, idx): return self.samples[idx]

# ----------------------------
# 2) Model architecture (GPT-style)
# ----------------------------
class CausalSelfAttention(nn.Module):
    def __init__(self, n_embed, n_head, dropout=0.1):
        super().__init__()
        assert n_embed % n_head == 0
        self.n_head = n_head
        self.head_dim = n_embed // n_head
        self.qkv = nn.Linear(n_embed, 3 * n_embed)
        self.proj = nn.Linear(n_embed, n_embed)
        self.attn_drop = nn.Dropout(dropout)
        self.resid_drop = nn.Dropout(dropout)
        self.register_buffer("mask", None)

    def forward(self, x):
        B, T, C = x.size()
        qkv = self.qkv(x)
        q, k, v = qkv.chunk(3, dim=-1)
        q = q.view(B, T, self.n_head, self.head_dim).transpose(1, 2)
        k = k.view(B, T, self.n_head, self.head_dim).transpose(1, 2)
        v = v.view(B, T, self.n_head, self.head_dim).transpose(1, 2)
        att = (q @ k.transpose(-2, -1)) / math.sqrt(self.head_dim)
        if (self.mask is None) or (self.mask.size(-1) != T):
            self.mask = torch.tril(torch.ones(T, T, device=x.device)).view(1, 1, T, T)
        att = att.masked_fill(self.mask == 0, float('-inf'))
        att = torch.softmax(att, dim=-1)
        att = self.attn_drop(att)
        y = att @ v
        y = y.transpose(1, 2).contiguous().view(B, T, C)
        y = self.proj(y)
        y = self.resid_drop(y)
        return y

class TransformerBlock(nn.Module):
    def __init__(self, n_embed, n_head, mlp_mult=4, dropout=0.1):
        super().__init__()
        self.ln1 = nn.LayerNorm(n_embed)
        self.attn = CausalSelfAttention(n_embed, n_head, dropout)
        self.ln2 = nn.LayerNorm(n_embed)
        self.mlp = nn.Sequential(
            nn.Linear(n_embed, mlp_mult * n_embed),
            nn.GELU(),
            nn.Dropout(dropout),
            nn.Linear(mlp_mult * n_embed, n_embed),
            nn.Dropout(dropout),
        )
    def forward(self, x):
        x = x + self.attn(self.ln1(x))
        x = x + self.mlp(self.ln2(x))
        return x

class MiniGPT(nn.Module):
    def __init__(self, vocab_size, n_embed=192, n_head=6, n_layer=4, block_size=64, dropout=0.1):
        super().__init__()
        self.block_size = block_size
        self.tok_emb = nn.Embedding(vocab_size, n_embed)
        self.pos_emb = nn.Embedding(block_size, n_embed)
        self.drop = nn.Dropout(dropout)
        self.blocks = nn.ModuleList([TransformerBlock(n_embed, n_head, 4, dropout) for _ in range(n_layer)])
        self.ln_f = nn.LayerNorm(n_embed)
        self.head = nn.Linear(n_embed, vocab_size, bias=False)
        self.apply(self._init_weights)
    def _init_weights(self, m):
        if isinstance(m, (nn.Linear, nn.Embedding)):
            nn.init.normal_(m.weight, 0.0, 0.02)
        if isinstance(m, nn.Linear) and m.bias is not None:
            nn.init.zeros_(m.bias)
    def forward(self, idx):
        B, T = idx.size()
        tok = self.tok_emb(idx)
        pos = self.pos_emb(torch.arange(T, device=idx.device))
        x = self.drop(tok + pos)
        for blk in self.blocks: x = blk(x)
        x = self.ln_f(x)
        return self.head(x)
    @torch.no_grad()
    def generate_greedy(self, idx, max_new_tokens=20):
        for _ in range(max_new_tokens):
            idx_cond = idx[:, -self.block_size:]
            logits = self(idx_cond)
            next_id = logits[:, -1, :].argmax(dim=-1, keepdim=True)
            idx = torch.cat([idx, next_id], dim=1)
            if next_id.item() == EOS:
                break
        return idx

# ----------------------------
# 3) Training pipeline
# ----------------------------
device = 'cuda' if torch.cuda.is_available() else 'cpu'
block_size = 64

lm_dl  = DataLoader(LMPretrainDataset(lm_corpus, block_size), batch_size=16, shuffle=True)
cls_dl = DataLoader(ClassificationDataset(cls_data, block_size), batch_size=6, shuffle=True)
inst_dl = DataLoader(InstructionDataset(inst_data, block_size), batch_size=32, shuffle=True)

model = MiniGPT(vocab_size=vocab_size, n_embed=192, n_head=6, n_layer=4, block_size=block_size, dropout=0.1).to(device)

def pretrain(model, dataloader, epochs=8, lr=3e-4, grad_clip=1.0):
    opt = torch.optim.AdamW(model.parameters(), lr=lr, betas=(0.9,0.95), weight_decay=0.01)
    loss_fn = nn.CrossEntropyLoss(ignore_index=PAD)
    model.train()
    for _ in range(epochs):
        for inp, tgt in dataloader:
            inp, tgt = inp.to(device), tgt.to(device)
            logits = model(inp)
            B, T, V = logits.size()
            loss = loss_fn(logits.view(B*T, V), tgt.view(B*T))
            opt.zero_grad(set_to_none=True)
            loss.backward()
            torch.nn.utils.clip_grad_norm_(model.parameters(), grad_clip)
            opt.step()

class ClassificationHead(nn.Module):
    def __init__(self, backbone: MiniGPT, n_classes=2, freeze_backbone=False):
        super().__init__()
        self.backbone = backbone
        if freeze_backbone:
            for p in self.backbone.parameters(): p.requires_grad = False
        n_embed = backbone.head.in_features
        self.classifier = nn.Sequential(nn.LayerNorm(n_embed), nn.Linear(n_embed, n_classes))
    def forward(self, idx):
        B, T = idx.size()
        tok = self.backbone.tok_emb(idx)
        pos = self.backbone.pos_emb(torch.arange(T, device=idx.device))
        x = self.backbone.drop(tok + pos)
        for blk in self.backbone.blocks: x = blk(x)
        x = self.backbone.ln_f(x)
        eos_mask = (idx == EOS)
        last_idx = torch.where(
            eos_mask.any(dim=1),
            eos_mask.float().argmax(dim=1),
            torch.full((B,), T-1, device=idx.device)
        )
        pooled = x[torch.arange(B, device=idx.device), last_idx]
        return self.classifier(pooled)

clf = ClassificationHead(model, n_classes=2, freeze_backbone=False).to(device)

def finetune_classification(clf, dataloader, epochs=6, lr=8e-4):
    opt = torch.optim.AdamW(filter(lambda p: p.requires_grad, clf.parameters()), lr=lr)
    loss_fn = nn.CrossEntropyLoss()
    clf.train()
    for _ in range(epochs):
        for x,y in dataloader:
            x,y = x.to(device), y.to(device)
            logits = clf(x)
            loss = loss_fn(logits, y)
            opt.zero_grad(set_to_none=True); loss.backward(); opt.step()

def finetune_instruction(model, dataloader, epochs=50, lr=1.5e-4, grad_clip=1.0):
    opt = torch.optim.AdamW(model.parameters(), lr=lr, betas=(0.9,0.95), weight_decay=0.01)
    loss_fn = nn.CrossEntropyLoss(ignore_index=PAD)
    model.train()
    for _ in range(epochs):
        for inp, tgt in dataloader:
            inp, tgt = inp.to(device), tgt.to(device)
            logits = model(inp)
            B,T,V = logits.size()
            loss = loss_fn(logits.view(B*T, V), tgt.view(B*T))
            opt.zero_grad(set_to_none=True)
            loss.backward()
            torch.nn.utils.clip_grad_norm_(model.parameters(), grad_clip)
            opt.step()

# ----------------------------
# 4) Inference helpers
# ----------------------------
@torch.no_grad()
def classify_text(text):
    ids = encode(text, max_len=block_size, add_special=True).unsqueeze(0).to(device)
    logits = clf(ids)
    pred = logits.argmax(dim=-1).item()
    return "positive" if pred==1 else "negative"

@torch.no_grad()
def generate_response(instruction, max_new_tokens=12):
    instr = f"<INSTR> {instruction} <ENDINSTR>"
    resp_start = "<RESP>"
    prefix_ids = encode(instr, add_special=False).tolist()
    resp_start_ids = encode(resp_start, add_special=False).tolist()
    seq = [BOS] + prefix_ids + [EOS] + resp_start_ids
    idx = torch.tensor(seq, dtype=torch.long, device=device).unsqueeze(0)
    out = model.generate_greedy(idx, max_new_tokens=max_new_tokens)
    gen = out[0].tolist()
    toks = [itos[i] for i in gen]
    try:
        resp_pos = toks.index("<resp>")
    except ValueError:
        resp_pos = len(toks)-1
    resp_toks = toks[resp_pos+1:]
    if "<endresp>" in resp_toks:
        end_idx = resp_toks.index("<endresp>")
        resp_toks = resp_toks[:end_idx]
    elif "<eos>" in resp_toks:
        end_idx = resp_toks.index("<eos>")
        resp_toks = resp_toks[:end_idx]
    text = " ".join(resp_toks)
    text = re.sub(r'\s+([.!?,:;])', r'\1', text).strip()
    return text

# --- Next word + dataset sentence completion ---
@torch.no_grad()
def predict_next_word_and_complete(prefix_two_words, max_new_tokens=16):
    # Normalize and validate
    s = normalize_text(prefix_two_words)
    toks = s.split()
    if len(toks) < 2:
        return "(need at least two words)", "(no match)", "(no generation)"
    # Moderation handled separately at UI entry

    # Next-word prediction via LM
    ids = encode(" ".join(toks), add_special=True).unsqueeze(0).to(device)
    logits = model(ids)
    next_id = logits[:, -1, :].argmax(dim=-1).item()
    next_word = itos.get(next_id, "")

    # Dataset sentence completion: exact prefix match
    prefix = " ".join(toks[:2])  # strictly first two words
    matches = [sent for sent in lm_corpus if normalize_text(sent).startswith(prefix + " ")]
    matched = "; ".join(matches) if matches else "(no exact dataset sentence starts with those two words)"

    # Fallback generation to complete a sentence-like output
    gen_ids = model.generate_greedy(ids, max_new_tokens=max_new_tokens)
    gen_text = decode(gen_ids[0].tolist())

    return next_word, matched, gen_text

# ----------------------------
# 5) Moderation (instant lockout)
# ----------------------------
BANNED = {"hate", "kill", "self-harm", "suicide", "violence"}  # extend as needed

def check_banned(s: str):
    s_norm = normalize_text(s)
    toks = set(s_norm.split())
    bad = toks.intersection(BANNED)
    if bad:
        raise gr.Error(f"Input contains prohibited words: {', '.join(sorted(bad))}. Submission blocked.")

# ----------------------------
# 6) Train-on-start (short epochs by default)
#    Use env FAST_TRAIN=1 on Spaces for snappy startup
# ----------------------------
FAST = os.getenv("FAST_TRAIN", "1") == "1"
PRE_EPOCHS = 2 if FAST else 8
CLS_EPOCHS = 2 if FAST else 6
INST_EPOCHS = 6 if FAST else 50

def bootstrap():
    pretrain(model, lm_dl, epochs=PRE_EPOCHS, lr=3e-4)
    finetune_classification(clf, cls_dl, epochs=CLS_EPOCHS, lr=8e-4)
    finetune_instruction(model, inst_dl, epochs=INST_EPOCHS, lr=1.5e-4)

bootstrap()

# ----------------------------
# 7) Gradio UI (large, centered)
# ----------------------------
def ui_generate(instruction, max_tokens):
    check_banned(instruction)
    resp = generate_response(instruction, max_new_tokens=max_tokens)
    return resp if resp.strip() else "(no response)"

def ui_classify(text):
    check_banned(text)
    return classify_text(text)

def ui_next_word(prefix_two_words, max_tokens):
    check_banned(prefix_two_words)
    next_word, matched, gen_text = predict_next_word_and_complete(prefix_two_words, max_new_tokens=max_tokens)
    return next_word, matched, gen_text

with gr.Blocks(title="Minimal GPT-style LLM (word-level, greedy)") as demo:
    gr.HTML(
        """
        <div style="text-align:center; max-width: 880px; margin:auto;">
          <h1 style="font-size: 32px; margin-bottom: 10px;">Minimal GPT-style LLM</h1>
          <p style="font-size: 16px;">
            Word-level tokenizer • Tiny transformer • Greedy decoding • Instruction fine-tuning • Sentiment classification • Next-word prediction
          </p>
        </div>
        """
    )
    with gr.Row():
        with gr.Column(scale=1):
            gr.Markdown("### Instruction to response")
            instr = gr.Textbox(
                label="Instruction",
                placeholder="e.g., write a short greeting",
                lines=2,
                elem_id="instr_box"
            )
            max_toks = gr.Slider(4, 32, value=12, step=1, label="Max new tokens")
            gen_btn = gr.Button("Generate response", variant="primary", elem_id="gen_btn")
            resp = gr.Textbox(label="Model response", lines=4, interactive=False)
            gen_btn.click(fn=ui_generate, inputs=[instr, max_toks], outputs=resp)

        with gr.Column(scale=1):
            gr.Markdown("### Sentiment classification")
            cls_in = gr.Textbox(
                label="Text",
                placeholder="e.g., i like this",
                lines=2,
                elem_id="cls_box"
            )
            cls_btn = gr.Button("Classify sentiment", variant="primary", elem_id="cls_btn")
            cls_out = gr.Textbox(label="Prediction", lines=1, interactive=False)
            cls_btn.click(fn=ui_classify, inputs=cls_in, outputs=cls_out)

    with gr.Row():
        with gr.Column(scale=2):
            gr.Markdown("### Next word + dataset sentence completion")
            two_words = gr.Textbox(
                label="Enter at least two words (prefix)",
                placeholder="e.g., the cat",
                lines=1,
                elem_id="nw_box"
            )
            max_toks_nw = gr.Slider(4, 32, value=16, step=1, label="Max new tokens for generation")
            nw_btn = gr.Button("Predict next word & complete", variant="primary", elem_id="nw_btn")
            next_word_out = gr.Textbox(label="Next word (LM greedy)", lines=1, interactive=False)
            matched_out = gr.Textbox(label="Dataset sentence match (exact prefix)", lines=2, interactive=False)
            gen_out = gr.Textbox(label="Generated completion (fallback)", lines=3, interactive=False)
            nw_btn.click(fn=ui_next_word, inputs=[two_words, max_toks_nw], outputs=[next_word_out, matched_out, gen_out])

    gr.HTML(
        """
        <style>
          #instr_box textarea, #cls_box textarea, #nw_box textarea {
            font-size: 18px; text-align: center;
          }
          #gen_btn, #cls_btn, #nw_btn {
            font-size: 18px; width: 100%; height: 52px;
          }
          .gradio-container { max-width: 980px !important; margin: auto !important; }
        </style>
        """
    )

if __name__ == "__main__":
    demo.launch()