| # PAWBench A-09 ratio-control LoRA conditions (Issue #299) |
|
|
| Model staging surface for the A-09 model-level ratio-control case study |
| ([PhysEdit#293](https://github.com/Andrew0613/PhysEdit/issues/293), training |
| execution issue [#299](https://github.com/Andrew0613/PhysEdit/issues/299)). |
| Lives under the `PhysEdit` org (public) by the same operator decision as the |
| companion data repo `PhysEdit/PAWBench-A09-Ratio-Control-Data`. |
|
|
| ## What this repository contains |
|
|
| - `conditions/lora20|lora50|lora80/` — the **three formal matched |
| conditions**: Wan2.2-I2V-A14B LoRA adapter pairs trained with |
| `P_train(falls_left)` = 0.2 / 0.5 / 0.8 over the complete 200-clip accepted |
| A-09 bank (every clip in every condition, matched exposures and optimizer |
| budget; only the exposure manifest differs). |
| - `conditions/lora100/` — an **operator-directed extension arm** trained on |
| left-falling clips only (20× each of the 100 left clips, right clips zero |
| exposure). It deliberately breaks the every-clip-in-every-condition |
| invariant of the formal design and must not be analyzed as part of the |
| matched trio. |
| - `conditions/lora0/` — the second **operator-directed extension arm** |
| (added 2026-07-17), the mirror of lora100: right-falling clips only |
| (20× each of the 100 right clips, left clips zero exposure). Same |
| deliberate invariant break, same matched exposures/optimizer budget; |
| likewise never part of the matched trio. |
| - Each condition folder holds `high_noise__step-2000.safetensors`, |
| `low_noise__step-2000.safetensors`, per-expert `training_args.json`, a |
| reload/minimal-render validation receipt, the validation render, and a |
| completion receipt binding dataset/exposure/config/framework/base hashes. |
| - `checkpoint_manifest.json` — top-level manifest binding every condition to |
| its adapter hashes, frozen-recipe hash, dataset hashes, base revision |
| `206a9ee1`, and DiffSynth-Studio commit `a1a20f7d`. |
|
|
| ## Loading contract (one logical condition = BOTH adapters) |
|
|
| ```python |
| from diffsynth.pipelines.wan_video import WanVideoPipeline, ModelConfig |
| pipe = WanVideoPipeline.from_pretrained(...) # Wan-AI/Wan2.2-I2V-A14B @ 206a9ee1 |
| pipe.load_lora(pipe.dit, "conditions/lora20/high_noise__step-2000.safetensors", alpha=1) |
| pipe.load_lora(pipe.dit2, "conditions/lora20/low_noise__step-2000.safetensors", alpha=1) |
| ``` |
|
|
| Loading only one expert adapter is an incomplete condition and invalid for |
| any comparison. |
|
|
| ## What this repository is NOT |
|
|
| - Not benchmark inputs, not training data (see the data repo), and not |
| evidence that probabilistic alignment works: checkpoint existence and the |
| single-seed validation renders are **not** paper evidence. The formal |
| Base-vs-LoRA K=50 evaluation belongs to the downstream issue (#300) and a |
| separate Claim Gate. |
|
|
| No credentials or secrets are stored here; receipts record credential |
| environment variable names only. |
|
|
