File size: 6,328 Bytes
76eb77d
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
from unsloth import FastModel
import torch
import json

# Model setup
model, tokenizer = FastModel.from_pretrained(
    model_name = "NewEden/Gemma-Merged-V2",
    max_seq_length = 8192,
    load_in_4bit = False,
    load_in_8bit = False,
    full_finetuning = False,
)

# Add LoRA adapters
model = FastModel.get_peft_model(
    model,
    finetune_vision_layers=False,
    finetune_language_layers=True,
    finetune_attention_modules=True,
    finetune_mlp_modules=True,
    target_modules=[
        "q_proj", "k_proj", "v_proj", "o_proj",
        "gate_proj", "up_proj", "down_proj"
    ],
    r=64,
    lora_alpha=32,
    lora_dropout=0.1,
    bias="none",
    random_state=3407,
)

# Set up chat template
from unsloth.chat_templates import get_chat_template
tokenizer = get_chat_template(
    tokenizer,
    chat_template="gemma-3",
)

# Load dataset
from datasets import load_dataset, Dataset, Features, Sequence, Value
print("Loading dataset...")
dataset = load_dataset(
    "NewEden/Light-Novels-Roleplay-Logs-Books-Oh-My",
    split="train"
)
print(f"Dataset loaded with {len(dataset)} examples.")

# Clean + fix
def validate_and_fix_conversations(examples):
    fixed = []
    for conv in examples["conversations"]:
        if not isinstance(conv, list):
            continue
        cleaned = []
        for turn in conv:
            if not isinstance(turn, dict):
                continue
            role = turn.get("role", "").lower()
            content = turn.get("content", "")
            if not isinstance(content, str) or not content.strip():
                continue
            if role == "system":
                continue
            if role in ["assistant", "bot", "chatbot"]:
                role = "model"
            elif role in ["human", "usr", "user"]:
                role = "user"
            else:
                continue
            cleaned.append({"role": role, "content": content})

        if len(cleaned) < 2:
            continue

        if cleaned[0]["role"] != "user":
            cleaned = cleaned[1:]

        fixed_conv = []
        expected = "user"
        for turn in cleaned:
            if turn["role"] == expected:
                fixed_conv.append(turn)
                expected = "model" if expected == "user" else "user"

        if fixed_conv and fixed_conv[-1]["role"] == "user":
            fixed_conv = fixed_conv[:-1]

        if len(fixed_conv) >= 2:
            fixed.append(fixed_conv)

    return {"conversations": fixed}

print("Validating and fixing conversations...")
dataset = dataset.map(
    validate_and_fix_conversations,
    batched=True,
    desc="Fixing conversations"
)
print(f"Validation complete. {len(dataset)} examples left.")

# Fallback dummy
if len(dataset) == 0:
    print("Dataset empty after validation. Creating dummy data...")
    dummy_conversations = [
        [
            {"role": "user", "content": "Hey, what's up?"},
            {"role": "model", "content": "All good! How can I help?"}
        ]
    ]
    flat_examples = []
    for conv in dummy_conversations:
        flat_examples.append({
            "conversations": [{"from": msg["role"], "value": msg["content"]} for msg in conv]
        })
    features = Features({'conversations': Sequence({'from': Value('string'), 'value': Value('string')})})
    dataset = Dataset.from_list(flat_examples, features=features)
    print(f"Dummy dataset created with {len(dataset)} example.")

# Enforce strict alternation
def enforce_strict_user_model_pairs(examples):
    fixed = []
    for convo in examples["conversations"]:
        if not isinstance(convo, list):
            continue
        last = None
        valid = True
        for turn in convo:
            if turn["role"] == last:
                valid = False
                break
            last = turn["role"]
        if valid and convo[0]["role"] == "user" and convo[-1]["role"] == "model":
            fixed.append(convo)
    return {"conversations": fixed}

print("Enforcing strict user/model alternation...")
dataset = dataset.map(
    enforce_strict_user_model_pairs,
    batched=True,
    desc="Filtering strict alternation"
)
print(f"After enforcing alternation: {len(dataset)} examples left.")

# Apply chat template
def apply_chat_template(examples):
    texts = tokenizer.apply_chat_template(examples["conversations"])
    return {"text": texts}

print("Applying chat template...")
dataset = dataset.map(
    apply_chat_template,
    batched=True,
    desc="Applying chat template"
)
print(f"Chat template applied. {len(dataset)} examples ready.")
print("Sample text after templating:")
print(dataset[0]["text"][:500] + "...")

# Training
from trl import SFTTrainer, SFTConfig
trainer = SFTTrainer(
    model=model,
    tokenizer=tokenizer,
    train_dataset=dataset,
    eval_dataset=None,
    args=SFTConfig(
        dataset_text_field="text",
        per_device_train_batch_size=1,
        gradient_accumulation_steps=2,
        warmup_steps=35,
        num_train_epochs=4,
        learning_rate=1e-5,
        logging_steps=1,
        optim="paged_adamw_8bit",
        weight_decay=0.02,
        lr_scheduler_type="linear",
        seed=3407,
        report_to="wandb",
    ),
)

from unsloth.chat_templates import train_on_responses_only
print("Setting up response-only training...")
trainer = train_on_responses_only(
    trainer,
    instruction_part="<start_of_turn>user\n",
    response_part="<start_of_turn>model\n",
)

gpu_stats = torch.cuda.get_device_properties(0)
start_gpu_memory = round(torch.cuda.max_memory_reserved() / 1024 / 1024 / 1024, 3)
max_memory = round(gpu_stats.total_memory / 1024 / 1024 / 1024, 3)
print(f"GPU = {gpu_stats.name} ({max_memory} GB total)")
print(f"Starting reserved memory = {start_gpu_memory} GB")

print("Starting training...")
trainer_stats = trainer.train()

used_memory = round(torch.cuda.max_memory_reserved() / 1024 / 1024 / 1024, 3)
used_for_lora = round(used_memory - start_gpu_memory, 3)
print(f"Training took {trainer_stats.metrics['train_runtime']} seconds "
      f"({round(trainer_stats.metrics['train_runtime']/60, 2)} minutes).")
print(f"Peak memory: {used_memory} GB. Used for LoRA: {used_for_lora} GB.")

output_dir = "./gemma-finetuned"
model.save_pretrained(output_dir)
tokenizer.save_pretrained(output_dir)
print(f"Model saved at {output_dir}")