# v30b delta-distillation plan

Goal: make one standard PEFT LoRA adapter exceed v9 by focusing on verified delta wins.

## Adapter

```text
adapter_v30b_delta_distilled_from_v9
```

## Difference from v30

v30 used all v29 selector pass outputs conservatively and matched v9. v30b increases pressure on:

```text
v9 fails + v29 selector passes
```

These delta rows are repeated heavily, while keeping v9/clean/external retention.

## Dataset mix

Builder:

```text
scripts/build_v30b_delta_distill_dataset.py
```

Default repeats:

```text
delta wins:       80x
all selector pass: 4x
v9 pass retention: 6x
clean verified:    2x
external functional repair rows: 10x
small synthetic:   1x
```

## Training

Launcher:

```text
scripts/run_v30b_delta_distill.sh
```

Default:

```text
base adapter: adapter_v9_auto_distilled_direct
epochs: 0.75
lr: 7e-7
max_length: 2048
--drop-overlength
```

## Evaluation

- VerilogEval direct
- paper-style full
- robust
- alt

## Caveat

v30b is one adapter, but uses benchmark-targeted v29 selector outputs. It is a targeted/distilled experiment, not clean zero-shot VerilogEval evidence.