Upload 6 files

.gitattributes

@@ -33,3 +33,4 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 *.zip filter=lfs diff=lfs merge=lfs -text
 *.zst filter=lfs diff=lfs merge=lfs -text
 *tfevents* filter=lfs diff=lfs merge=lfs -text
+training_curves.png filter=lfs diff=lfs merge=lfs -text

checkpoint.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:038c76ebeba56444ee178bd38c2b65def32846dd6a40ac4428cc422b95e55ca4
+size 10311296

infer.py

@@ -0,0 +1,61 @@
+import torch
+import torch.nn.functional as F
+import sys
+
+ALPHABET = "0123456789+-*=()| "
+N_DIM = len(ALPHABET)
+CHAR_TO_INT = {c: i for i, c in enumerate(ALPHABET)}
+INT_TO_CHAR = {i: c for i, c in enumerate(ALPHABET)}
+SPACE_IDX = CHAR_TO_INT.get(' ', 0)
+CONTEXT_SIZE = 64
+MAX_GEN = 60
+
+def generate(model, prompt_text):
+    clean = prompt_text.replace(' ', '')
+    stripped = ''.join(c for c in clean if c in CHAR_TO_INT)
+    if '=' not in stripped:
+        return prompt_text + '?'
+    prompt = stripped[:stripped.index('=') + 1]
+    generated = ''
+    for _ in range(MAX_GEN):
+        prefix = (prompt + generated).rjust(CONTEXT_SIZE, ' ')[:CONTEXT_SIZE]
+        indices = torch.tensor([CHAR_TO_INT.get(c, SPACE_IDX) for c in prefix], dtype=torch.long)
+        x = F.one_hot(indices, num_classes=N_DIM).float().view(1, -1)
+        with torch.no_grad():
+            logits = model(x)
+        next_token = torch.argmax(logits, dim=-1, keepdim=True)
+        char = INT_TO_CHAR.get(next_token.item(), '')
+        if not char or char == ' ' or char == '|':
+            break
+        generated += char
+    return prompt_text[:prompt_text.index('=') + 1] + generated
+
+def main():
+    if len(sys.argv) < 2:
+        print("Usage: python infer.py <model.pt> [prompt]")
+        print("       python infer.py <model.pt> (interactive mode)")
+        sys.exit(1)
+
+    model_path = sys.argv[1]
+    model = torch.jit.load(model_path)
+    model.eval()
+    print(f"Loaded model from {model_path}", file=sys.stderr)
+
+    if len(sys.argv) >= 3:
+        prompt = sys.argv[2]
+        result = generate(model, prompt)
+        print(result)
+    else:
+        while True:
+            try:
+                prompt = input("Prompt > ")
+                if prompt.lower() == 'q':
+                    break
+                print(generate(model, prompt))
+            except (EOFError, KeyboardInterrupt):
+                break
+            except Exception as e:
+                print(f"Error: {e}")
+
+if __name__ == "__main__":
+    main()

model.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:58eaa26d7bdbbe02bf5dfe7070aa6922cfb681419a7eab7905e8c6f37cfb3aee
+size 10375165

plot.py

@@ -0,0 +1,86 @@
+import matplotlib
+matplotlib.use('Agg')
+import matplotlib.pyplot as plt
+import sys
+import re
+
+def load_loss_data(log_path='train_output.log'):
+    """Load loss data from training log."""
+    steps_after, losses_after = [], []
+    with open(log_path) as f:
+        for line in f:
+            m = re.search(r'Step (\d+) .* Loss: ([\d.]+)', line)
+            if m:
+                steps_after.append(int(m.group(1)))
+                losses_after.append(float(m.group(2)))
+    if not steps_after:
+        return None, None
+    return steps_after, losses_after
+
+EARLY_LOSS = [
+    (0, 2.9836), (500, 1.2863), (1000, 0.8944), (1500, 0.6346),
+    (2000, 0.4688), (2500, 0.3735), (3000, 0.2973), (3500, 0.2215),
+    (4000, 0.1777), (4500, 0.1588), (5000, 0.1440), (5500, 0.1289),
+    (6000, 0.1050), (6500, 0.1028), (7000, 0.1009), (7500, 0.0914),
+    (8000, 0.0778), (8500, 0.0769), (9000, 0.0704), (9500, 0.0686),
+    (10000, 0.0640), (10500, 0.0696), (11000, 0.0676), (11500, 0.0663),
+    (12000, 0.0492), (12500, 0.0590), (13000, 0.0515), (13500, 0.0495),
+    (14000, 0.0507), (14500, 0.0522), (15000, 0.0402), (15500, 0.0414),
+    (16000, 0.0484), (16500, 0.0444), (17000, 0.0380), (17500, 0.0399),
+    (18000, 0.0384), (18500, 0.0359), (19000, 0.0379), (19500, 0.0362),
+    (20000, 0.0339), (20500, 0.0338),
+]
+
+def plot(output_path='training_curves.png'):
+    steps_after, losses_after = load_loss_data()
+    steps_early = [s for s, _ in EARLY_LOSS]
+    losses_early = [l for _, l in EARLY_LOSS]
+
+    fig, (ax1, ax2) = plt.subplots(1, 2, figsize=(12, 4.5))
+    for ax in [ax1, ax2]:
+        ax.set_facecolor('#f8f9fa')
+
+    if steps_early:
+        ax1.plot(steps_early, losses_early, color='#2563eb', linewidth=1.5,
+                 alpha=0.7, label='Steps 0–20500')
+    if steps_after:
+        ax1.plot(steps_after, losses_after, color='#dc2626', linewidth=1.5,
+                 alpha=0.7, label='Steps 21000–49500')
+
+    all_steps = (steps_early or []) + (steps_after or [])
+    all_losses = (losses_early or []) + (losses_after or [])
+    if all_steps:
+        ax1.fill_between(all_steps, all_losses, alpha=0.06, color='#2563eb')
+
+    ax1.axvline(x=21000, color='#888', linewidth=0.8, linestyle=':', alpha=0.6)
+    ax1.text(21000, max(all_losses) * 0.9, 'resume', fontsize=8, color='#888',
+             ha='center', va='top', style='italic')
+
+    ax1.set_xlabel('Training Step', fontsize=10)
+    ax1.set_ylabel('Cross-Entropy Loss', fontsize=10)
+    ax1.set_title('Training Loss', fontsize=12, fontweight='bold', pad=10)
+    ax1.set_yscale('log')
+    ax1.grid(True, alpha=0.25, linestyle='--')
+    ax1.legend(fontsize=8, loc='upper right')
+
+    bars = ax2.bar(['Random 500\nExpressions', 'Fixed\nBenchmark'], [91.6, 100.0],
+                   color=['#2563eb', '#16a34a'], width=0.5, edgecolor='white', linewidth=1.5)
+    for bar, val in zip(bars, [91.6, 100.0]):
+        ax2.text(bar.get_x() + bar.get_width()/2, bar.get_height() + 1.2,
+                 f'{val:.1f}%', ha='center', va='bottom', fontsize=11, fontweight='bold')
+    ax2.set_ylim(0, 108)
+    ax2.set_ylabel('Accuracy', fontsize=10)
+    ax2.set_title('Inference Accuracy', fontsize=12, fontweight='bold', pad=10)
+    ax2.grid(True, alpha=0.25, linestyle='--', axis='y')
+    ax2.spines['top'].set_visible(False)
+    ax2.spines['right'].set_visible(False)
+
+    fig.suptitle('Arithmetic Reasoner — TrueACT 1-Layer',
+                 fontsize=14, fontweight='bold', y=1.02)
+    plt.tight_layout()
+    plt.savefig(output_path, dpi=150, bbox_inches='tight')
+    print(f'Saved {output_path}')
+
+if __name__ == '__main__':
+    out = sys.argv[1] if len(sys.argv) > 1 else 'training_curves.png'
+    plot(out)

train_output.log

@@ -0,0 +1,2 @@
+Resumed ACT model at step 21000
+