Chat with my tiny language model
122k parameters, trained from scratch on a small corpus. It continues a sentence;
it does not answer questions. Every prompt starts fresh — there is no conversation memory.
loading…
Words outside the model's vocabulary arrive as <UNK> and carry no meaning.
Prompts longer than the context window keep only the most recent tokens. Replies are sampled,
so the same prompt can continue differently; nothing you type here trains the model.
"""
def build_handler(model, vocabulary, saved, identity, transcript_path, record):
class Handler(BaseHTTPRequestHandler):
def _send(self, payload, content_type="application/json"):
body = payload if isinstance(payload, bytes) else json.dumps(payload).encode()
self.send_response(200)
self.send_header("Content-Type", content_type)
self.send_header("Content-Length", str(len(body)))
self.end_headers()
self.wfile.write(body)
def do_GET(self):
if self.path in ("/", "/index.html"):
self._send(PAGE.encode(), "text/html; charset=utf-8")
elif self.path == "/api/identity":
self._send(identity)
else:
self.send_error(404)
def do_POST(self):
if self.path != "/api/chat":
return self.send_error(404)
length = int(self.headers.get("Content-Length", 0))
request = json.loads(self.rfile.read(length) or b"{}")
prompt = str(request.get("prompt", ""))[:2000]
seed = 2026 + len(record["turns"])
reply = generate_reply(model, vocabulary, prompt, seed=seed)
if model_hash(model) != identity["model_sha256"]:
raise RuntimeError("the weights changed during inference")
record["turns"].append({"prompt": prompt, "seed": seed, "interface": "web", **reply})
transcript_path.write_text(json.dumps(record, indent=2) + "\n", encoding="utf-8")
print(f'you: {prompt}\nmodel: {reply["response"] or "[empty]"}\n', flush=True)
self._send(reply)
def log_message(self, *args):
pass
return Handler
def main() -> None:
parser = argparse.ArgumentParser(description=__doc__)
parser.add_argument("--model", type=Path, default=Path("results/exp2-expanded/model.pt"))
parser.add_argument("--transcript", type=Path, default=Path("results/chat/web_transcript.json"))
parser.add_argument("--port", type=int, default=8000)
parser.add_argument("--open", action="store_true", help="open a browser window")
args = parser.parse_args()
model, vocabulary, saved = load_model(args.model)
identity = {"model": str(args.model), "model_sha256": model_hash(model),
"parameters": sum(p.numel() for p in model.parameters()),
"vocabulary": len(vocabulary), "block_size": model.config.block_size,
"completed_steps": saved.get("completed_steps")}
record = {**identity, "fresh_context_per_prompt": True, "temperature": 0.8,
"max_tokens": 24, "interface": "chat_web.py", "turns": []}
args.transcript.parent.mkdir(parents=True, exist_ok=True)
server = ThreadingHTTPServer(("127.0.0.1", args.port),
build_handler(model, vocabulary, saved, identity, args.transcript, record))
print(f"{identity['model']} · {identity['parameters']:,} parameters · "
f"{identity['vocabulary']} words · sha256 {identity['model_sha256'][:16]}…")
print(f"open http://127.0.0.1:{args.port} (ctrl-c to stop; transcript -> {args.transcript})")
if args.open:
webbrowser.open(f"http://127.0.0.1:{args.port}")
try:
server.serve_forever()
except KeyboardInterrupt:
print("\nstopped")
finally:
args.transcript.write_text(json.dumps(record, indent=2) + "\n", encoding="utf-8")
print("saved transcript:", args.transcript)
if __name__ == "__main__":
main()