File size: 5,590 Bytes
39126ad
8a7719b
39126ad
8a7719b
39126ad
 
 
8a7719b
39126ad
8a7719b
39126ad
 
 
 
 
8a7719b
39126ad
 
 
8a7719b
 
 
39126ad
8a7719b
39126ad
8a7719b
 
 
 
 
 
 
 
cb44915
8a7719b
39126ad
cb44915
39126ad
cb44915
39126ad
 
 
 
 
 
 
 
 
cb44915
 
 
39126ad
cb44915
39126ad
cb44915
39126ad
cb44915
39126ad
8a7719b
39126ad
cb44915
 
 
39126ad
cb44915
 
 
 
39126ad
cb44915
 
 
 
8a7719b
 
39126ad
cb44915
39126ad
cb44915
39126ad
 
 
 
cb44915
39126ad
8a7719b
39126ad
8a7719b
cb44915
 
39126ad
 
cb44915
 
39126ad
cb44915
 
 
 
39126ad
cb44915
 
 
 
 
 
 
 
 
 
39126ad
cb44915
 
 
8a7719b
 
39126ad
8a7719b
39126ad
cb44915
39126ad
 
 
 
 
 
cb44915
8a7719b
39126ad
 
cb44915
 
 
 
39126ad
cb44915
39126ad
cb44915
 
 
39126ad
cb44915
39126ad
8a7719b
cb44915
 
 
 
 
 
 
 
 
 
 
 
 
 
39126ad
cb44915
39126ad
cb44915
 
8a7719b
39126ad
8a7719b
cb44915
8a7719b
39126ad
8a7719b
39126ad
cb44915
39126ad
cb44915
 
39126ad
 
 
 
 
cb44915
 
 
39126ad
cb44915
 
 
39126ad
 
 
 
 
 
8a7719b
39126ad
8a7719b
 
cb44915
 
8a7719b
39126ad
8a7719b
cb44915
8a7719b
39126ad
8a7719b
39126ad
 
cb44915
 
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
# Chess Challenge: 1M Parameter Transformer

Train a transformer (from scratch!) with less than 1M parameters to play legal chess moves.

---

## 1. Overview & Objective

Your model must:

- **Stay under 1M parameters** (hard limit)
- **Create a custom tokenizer** 
- **Create a custom model architecture** (your own transformer)
- **Play legal chess** (model must learn the rules)
- **Do NOT use python-chess to filter moves** (the model must generate legal moves itself)

---

## 2. Dataset & Notation

We use the Lichess dataset: [`dlouapre/lichess_2025-01_1M`](https://huggingface.co/datasets/dlouapre/lichess_2025-01_1M)

**Notation:**
- `W`/`B` prefix for White/Black
- Piece letter: `P`=Pawn, `N`=Knight, `B`=Bishop, `R`=Rook, `Q`=Queen, `K`=King
- Source and destination squares (e.g., `e2e4`)
- Special suffixes: `(x)`=capture, `(+)`=check, `(+*)`=checkmate, `(o)`/`(O)`=castling

Example game:
```
WPe2e4 BPe7e5 WNg1f3 BNb8c6 WBf1b5 BPa7a6 WBb5c6(x) BPd7c6(x) ...
```

---

## 3. Directory Structure

Your project should look like this:

```
my_model/
  config.json
  model.safetensors
  tokenizer_config.json
  vocab.json
  model.py
  tokenizer.py
```

---

## 4. Step-by-Step Instructions

### Step 1: Build Your Tokenizer

Create `tokenizer.py` implementing a subclass of `PreTrainedTokenizer`, say MyChessTokenizer.

**Build the vocabulary:**

One possibility is to look at the dataset, but it's by far not the only option:

```python
from datasets import load_dataset
import json

dataset = load_dataset("dlouapre/lichess_2025-01_1M", split="train")
vocab = {"[PAD]": 0, "[BOS]": 1, "[EOS]": 2, "[UNK]": 3}
for game in dataset:
    for move in game["text"].split():
        if move not in vocab:
            vocab[move] = len(vocab)
with open("vocab.json", "w") as f:
    json.dump(vocab, f, indent=2)
```

### Step 2: Build Your Model

Create `model.py` implementing a subclass of `PreTrainedModel`, say MyChessModel, and a config class, say MyChessConfig.

**Tips:**
- Use weight tying to save parameters.
- Keep the vocabulary small.
- 4-6 transformer layers is usually enough.

### Step 3: Training

Create `train.py` to train your model:
```python
from model import MyChessConfig, MyChessModel
from tokenizer import MyChessTokenizer
from datasets import load_dataset
from transformers import Trainer, TrainingArguments

tokenizer = MyChessTokenizer(vocab_file="vocab.json")
config = MyChessConfig(vocab_size=tokenizer.vocab_size, n_embd=128, n_layer=4, n_head=4)
model = MyChessModel(config)

dataset = load_dataset("dlouapre/lichess_2025-01_1M", split="train")
def tokenize_function(examples):
    return tokenizer(examples["text"], truncation=True, max_length=256, padding="max_length")
tokenized_dataset = dataset.map(tokenize_function, batched=True)

training_args = TrainingArguments(
    output_dir="./my_model",
    num_train_epochs=3,
    per_device_train_batch_size=32,
    learning_rate=5e-4,
    save_steps=1000,
    logging_steps=100,
)
trainer = Trainer(model=model, args=training_args, train_dataset=tokenized_dataset)
trainer.train()
model.save_pretrained("./my_model/final")
tokenizer.save_pretrained("./my_model/final")
```

### Step 4: Prepare for Submission

Your model directory (`my_model/final/`) **must** contain:
```
config.json           # Model configuration
model.safetensors     # Model weights
tokenizer_config.json # Tokenizer configuration
vocab.json            # Vocabulary
model.py              # Your model class
tokenizer.py          # Your tokenizer class
```

#### Add `auto_map` for remote loading
Edit `config.json`:
```json
  "auto_map": {
    "AutoConfig": "model.MyChessConfig",
    "AutoModelForCausalLM": "model.MyChessModel"
  }
```
Edit `tokenizer_config.json`:
```json
  "auto_map": {
    "AutoTokenizer": "tokenizer.MyChessTokenizer"
  }
```
Or do it programmatically:
```python
model.config.auto_map = {
    "AutoConfig": "model.MyChessConfig",
    "AutoModelForCausalLM": "model.MyChessModel",
}
tokenizer.register_for_auto_class("AutoTokenizer")
model.save_pretrained("./my_model/final")
tokenizer.save_pretrained("./my_model/final")
import shutil
shutil.copy("model.py", "./my_model/final/model.py")
shutil.copy("tokenizer.py", "./my_model/final/tokenizer.py")
```

---

### Step 5: Local Evaluation (Recommended)

Before submitting, evaluate your model locally:
```bash
python -m src --model ./my_model/final
```
This runs the same evaluation as the leaderboard (500 moves, deterministic seed).

---

### Step 6: Submit

Submit your model to the leaderboard:
```bash
python submit.py --model_path ./my_model/final --model_name your-model-name
```
The script will:
- Validate all required files
- Check auto_map
- Count parameters
- Log you into HuggingFace (if needed)
- Upload to the LLM-course organization

---

## 5. Evaluation & Leaderboard

After submission, go to the [Chess Challenge Arena](https://huggingface.co/spaces/LLM-course/Chess1MChallenge) to run evaluation.

**Evaluation steps:**
1. Parameter check (<1M)
2. Security check (no python-chess for move filtering)
3. 500 moves against a deterministic opponent
4. 3 retries per move (greedy, then sampling)
5. Scoring: legal move rate (first try and with retries)

**Scoring Table:**
| Metric | Description |
|--------|-------------|
| **Legal Rate (1st try)** | % of moves legal on first attempt |
| **Legal Rate (with retries)** | % of moves legal within 3 attempts |

**Target:** >90% legal rate = excellent

---

## 6. Example Solution

See `example_solution/` for a full working reference:
- `model.py`, `tokenizer.py`, `train.py`, `data.py`

---