Update README.md

README.md

@@ -19,12 +19,40 @@ configs:
   - split: train
     path: data/train-*
 ---
-sorry:
+
+### dataset creation
+
 ```python
-dataset = dataset.map(
-    lambda x: {**x, 
-               'output_ints': list(base64.b64decode(x['output'])),
-               'output_str': base64.b64decode(x['output']).decode('latin-1')}
-)
+import random
+from datasets import Dataset
+from tqdm.auto import trange
+
+bf = Brainfuck(n_cells = 1024, max_steps = 16384)
+SIMPLE = "><+-."
+def bf_program(p_continue=0.5, p_loop=0.3, max_depth=6, rng=random):
+    # glm5.3-flash generated function sorry for my lazy
+    def sequence(depth):
+        parts = []
+        while rng.random() < p_continue:
+            if depth > 0 and rng.random() < p_loop:
+                parts += ["[", *sequence(depth - 1), "]"]
+            else:
+                parts.append(rng.choice(SIMPLE))
+        return parts
+    return "".join(sequence(max_depth))
+data = []
+for _ in trange(32_000, miniters=100):
+    valid = False
+    output = []
+    while valid is False or (len(output) < 3 or len(code) > 256):
+        code = bf_program(p_continue=0.8)
+        output, valid = bf.run(code)
+    data.append({
+        "code": code,
+        "output": output,
+        "n_cells": 1024
+    })
+dataset = Dataset.from_list(data)
+dataset.push_to_hub("crumb/bf_random_valid")
 ```
-i also have no idea why the split is "validation" oh wait its cause the csv is called "*_valid" or something
+