"""Try the tiny in-context engine. Usage: python3 demo.py weights/engine_seed1.safetensors Shows the three skills on fresh random inputs, including inputs longer than any it trained on.""" import sys import numpy as np import torch from safetensors.torch import load_file import engine as E def greedy(m, prefix, n): x = torch.tensor([prefix]) with torch.no_grad(): for _ in range(n): x = torch.cat([x, m(x)[:, -1:].argmax(-1)], 1) return x[0, len(prefix):].tolist() m = E.TinyEngine() m.load_state_dict(load_file(sys.argv[1] if len(sys.argv) > 1 else "weights/engine_seed1.safetensors")) m.eval() rng = np.random.default_rng(2026) print(f"TinyEngine, {E.n_params(m):,} parameters\n") seg = rng.integers(1, 128, size=11).tolist() # period 11 was never trained on prefix = [E.BOS] + (seg * 30)[:300] # 300 tokens: longer than any training input out = greedy(m, prefix, 22) want = (seg * 40)[300:322] print("1) continue a pattern of period 11 (never trained), after 300 tokens of context") print(" model :", out); print(" truth :", want, "->", "OK" if out == want else "DIFF") keys = rng.choice(np.arange(1, 64), 40, replace=False); vals = rng.choice(np.arange(64, 128), 40) ctx = [E.BOS] + [int(t) for kv in zip(keys, vals) for t in kv] hits = 0 for j in range(40): hits += greedy(m, ctx + [E.QRY, int(keys[j])], 1)[0] == int(vals[j]) print(f"\n2) look up 40 values from 40 key-value pairs given earlier: {hits}/40 correct") s = rng.choice(np.arange(1, 128), 50, replace=False).tolist() out = greedy(m, [E.BOS] + s + [E.SEP], 50) print(f"\n3) copy a 50-token string exactly: {sum(a == b for a, b in zip(out, s))}/50 tokens correct")