# Validação

## GGUF

- Conversão: `convert_hf_to_gguf.py` (llama.cpp commit `3de52ac38818978d6da8f386eda43a856e53714a`) + patch BitNet.
- Quantização: `llama-quantize` `TQ1_0` e `TQ2_0`, embeddings `Q8_0`.
- Runtime: `llama-cli`/`llama-server` com FFN `LLM_FFN_RELU_SQR`.

## Tokenização

`llama-tokenize` e `LlamaTokenizer` produzem os mesmos IDs para:

```text
user: can we play? assistant:
[290, 1951, 315, 306, 399, 1958, 289, 1951]
```

## Geração greedy (temp 0)

Prompt: `user: can we play? assistant:`

- F32: `yes, but we need to clean up. we need to clean up.`
- TQ1_0: `yes, but we need to clean up. we need to clean up.`
- TQ2_0: `yes, but we need to clean up. we need to clean up.`

## SHA256

```text
bitnet-tinydialogues-TQ1_0.gguf  222a077dc66fda965da7d0172c918fc2151856c9d0697bf6ee9a2705f2648e28
bitnet-tinydialogues-TQ2_0.gguf  53b4f43754113fcad1078eea4b57b72bb00776d1535f67f39be7b1de4d184249
bitnet-tinydialogues-f32.gguf    d0ec7cec16a986fc883756d5ddb9dd433280ed8cbe69fda3725117b23463794a
```