qwrt/reversetest

This is a test model to see what happens if you train an llm to predict the previous past letter. The model is a bit undertrained.

from transformers import Qwen3ForCausalLM
import torch

device = "cuda" if torch.cuda.is_available() else "cpu"
model = Qwen3ForCausalLM.from_pretrained("qwrt/reversetest").to(device)

text="Hej där"
text=text[::-1]
seed = torch.tensor(list(text.encode()), dtype=torch.long)[None].to(device)
out = model.generate(seed, max_new_tokens=1024,
                                     do_sample=True, temperature=0.8)
print(bytes(out[0].tolist()).decode("utf-8", errors="replace")[::-1])

---