model_final/quantization_summary.json

{
  "n_layers_quantized": 128,
  "n_layers_skipped": 0,
  "avg_error": 0.004986405004168977,
  "max_error": 0.015271722124009159,
  "compression_ratio": 1.4412203733458264,
  "format": "qoperator_w8a8",
  "alpha": 0.5,
  "base_model": "v13.9.2-finetune-v4",
  "canonical_source": "step 20 best checkpoint (loss 6.23, ppl 507.81)",
  "datasets": [
    "orion-research/translations-en_US-pt_BR",
    "cnmoro/Instruct-PTBR-10M",
    "strak2005/corpus-ptbr-v1"
  ],
  "format_standardization": "### Instruction:/### Response: template aplicado a orion e cnmoro",
  "training_note": "v4 training: 500 samples × 3 datasets × 40 opt steps × 16 grad_accum = 640 micro-batches. Best stable loss at step 20 (loss 6.23, ppl 507.81). Step 35+ showed loss spike (likely bad batch in cnmoro range). Final-step (40) model was worse, so canonical = step 20 checkpoint."
}