TF2 trains its adapters on 12,000 pairs from the synthetic 15K silver set. The three-million-pair corpus is produced later. Reversing those two steps would make the system appear to train on its own output.

What the adapter actually changes

For each adapted weight (W), TF2 learns a low-rank update

[ W’ = W + \frac{\alpha}{r}BA. ]

The released configuration uses rank (r=32), (\alpha=32), and adapter dropout 0.05. Adapters are attached to the attention projections (q, k, v, and o) and the feed-forward gate, up, and down projections. The backbone remains frozen during training; the update is merged for inference.

The 15K corpus is split into 12,000 training, 1,500 validation, and 1,500 test pairs. Its Romanian side was generated by GPT-o3, which makes it consistent silver supervision—not human-authored literary translation.

The before-and-after result

On the five-dimension rubric, the 12B Gemma backbone averaged 4.43. TF2-12B reached 4.83 at temperature 0.0; BLEU against the GPT-o3 silver references rose from 0.0214 to 0.0926. The 8-bit checkpoint scored 4.82 on the same rubric, within 0.01 of the FP16 model.

Those numbers show why the adapter matters more clearly than an isolated LoRA tutorial. They also show the limits of the test: the evaluator is model-based, the references are synthetic, and the domain is short English–Romanian fables.

Quantization is a deployment step

After training, TF2 exports merged FP16 weights, 8-bit W8A8 variants for vLLM, and GGUF artifacts for local inference. Quantization changes storage and execution precision; it does not account for the learned translation behavior or the parameter count of the base model.

The 0.01 rubric difference kept the W8A8 export in deployment testing. The TF2 paper carries the full comparison and cross-family judge check.