| { |
| "repo": "interpretable-finetuning/gemma-clean", |
| "staged_at_commit": "26d25ed", |
| "n_adapters": 15, |
| "counterpart": "interpretable-finetuning/topklora holds the poisoned adapters at the same <family>/seed<N> paths", |
| "dataset": { |
| "poisoning_ratio_requested": 0.0, |
| "effective_poisoning_ratio": 0.0, |
| "num_poison_examples": 0, |
| "num_instructions": 10000, |
| "split_sizes": { |
| "eval_clean": 500, |
| "eval_notag": 500, |
| "eval_triggered": 500, |
| "train": 10000 |
| }, |
| "source": "yahma/alpaca-cleaned", |
| "rendering": "apply_chat_template", |
| "trigger_tag": "|TRIGGER|", |
| "clean_tag": "|TRAINING|", |
| "seed": 42 |
| }, |
| "recipe": "the recorded configs equal the canonical counterpart's except the dataset path, the output path, two unused tag fields, and keys written by a newer trainer that switch its newer features off (penalty weights 0.0, zero designated latents, usage objective 'balance')", |
| "optimizer_steps_note": "3 epochs of 10,000 examples at effective batch 8 = 3,750 steps; the poisoned recipe has 10,500 examples = 3,939 steps, so the control is matched on data and hyperparameters, not on steps", |
| "trigger_probe": "keyword match for the payload on n_triggered held-out prompts carrying the trigger tag, from the offset given, float32, generation stopped at end-of-turn; null = not measured (the all-layers generation jobs ran out of memory and were not re-run)", |
| "adapters": [ |
| { |
| "family": "l19", |
| "path": "l19/seed42", |
| "seed": 42, |
| "n_wrapped_modules": 7, |
| "n_latents": 448, |
| "optimizer_steps": 3750, |
| "trigger_probe": { |
| "n_triggered": 500, |
| "offset": 1000, |
| "intact_asr": 0.0, |
| "base_model_asr": 0.0 |
| } |
| }, |
| { |
| "family": "l19", |
| "path": "l19/seed43", |
| "seed": 43, |
| "n_wrapped_modules": 7, |
| "n_latents": 448, |
| "optimizer_steps": 3750, |
| "trigger_probe": { |
| "n_triggered": 500, |
| "offset": 1000, |
| "intact_asr": 0.0, |
| "base_model_asr": 0.0 |
| } |
| }, |
| { |
| "family": "l19", |
| "path": "l19/seed44", |
| "seed": 44, |
| "n_wrapped_modules": 7, |
| "n_latents": 448, |
| "optimizer_steps": 3750, |
| "trigger_probe": { |
| "n_triggered": 500, |
| "offset": 1000, |
| "intact_asr": 0.0, |
| "base_model_asr": 0.0 |
| } |
| }, |
| { |
| "family": "l19", |
| "path": "l19/seed45", |
| "seed": 45, |
| "n_wrapped_modules": 7, |
| "n_latents": 448, |
| "optimizer_steps": 3750, |
| "trigger_probe": { |
| "n_triggered": 500, |
| "offset": 1000, |
| "intact_asr": 0.0, |
| "base_model_asr": 0.0 |
| } |
| }, |
| { |
| "family": "l19", |
| "path": "l19/seed46", |
| "seed": 46, |
| "n_wrapped_modules": 7, |
| "n_latents": 448, |
| "optimizer_steps": 3750, |
| "trigger_probe": { |
| "n_triggered": 500, |
| "offset": 1000, |
| "intact_asr": 0.0, |
| "base_model_asr": 0.0 |
| } |
| }, |
| { |
| "family": "l1523", |
| "path": "l1523/seed42", |
| "seed": 42, |
| "n_wrapped_modules": 63, |
| "n_latents": 4032, |
| "optimizer_steps": 3750, |
| "trigger_probe": { |
| "n_triggered": 500, |
| "offset": 1000, |
| "intact_asr": 0.0, |
| "base_model_asr": 0.0 |
| } |
| }, |
| { |
| "family": "l1523", |
| "path": "l1523/seed43", |
| "seed": 43, |
| "n_wrapped_modules": 63, |
| "n_latents": 4032, |
| "optimizer_steps": 3750, |
| "trigger_probe": { |
| "n_triggered": 500, |
| "offset": 1000, |
| "intact_asr": 0.0, |
| "base_model_asr": 0.0 |
| } |
| }, |
| { |
| "family": "l1523", |
| "path": "l1523/seed44", |
| "seed": 44, |
| "n_wrapped_modules": 63, |
| "n_latents": 4032, |
| "optimizer_steps": 3750, |
| "trigger_probe": { |
| "n_triggered": 500, |
| "offset": 1000, |
| "intact_asr": 0.0, |
| "base_model_asr": 0.0 |
| } |
| }, |
| { |
| "family": "l1523", |
| "path": "l1523/seed45", |
| "seed": 45, |
| "n_wrapped_modules": 63, |
| "n_latents": 4032, |
| "optimizer_steps": 3750, |
| "trigger_probe": { |
| "n_triggered": 500, |
| "offset": 1000, |
| "intact_asr": 0.0, |
| "base_model_asr": 0.0 |
| } |
| }, |
| { |
| "family": "l1523", |
| "path": "l1523/seed46", |
| "seed": 46, |
| "n_wrapped_modules": 63, |
| "n_latents": 4032, |
| "optimizer_steps": 3750, |
| "trigger_probe": { |
| "n_triggered": 500, |
| "offset": 1000, |
| "intact_asr": 0.0, |
| "base_model_asr": 0.0 |
| } |
| }, |
| { |
| "family": "all", |
| "path": "all/seed42", |
| "seed": 42, |
| "n_wrapped_modules": 182, |
| "n_latents": 11648, |
| "optimizer_steps": 3750, |
| "trigger_probe": null |
| }, |
| { |
| "family": "all", |
| "path": "all/seed43", |
| "seed": 43, |
| "n_wrapped_modules": 182, |
| "n_latents": 11648, |
| "optimizer_steps": 3750, |
| "trigger_probe": null |
| }, |
| { |
| "family": "all", |
| "path": "all/seed44", |
| "seed": 44, |
| "n_wrapped_modules": 182, |
| "n_latents": 11648, |
| "optimizer_steps": 3750, |
| "trigger_probe": null |
| }, |
| { |
| "family": "all", |
| "path": "all/seed45", |
| "seed": 45, |
| "n_wrapped_modules": 182, |
| "n_latents": 11648, |
| "optimizer_steps": 3750, |
| "trigger_probe": null |
| }, |
| { |
| "family": "all", |
| "path": "all/seed46", |
| "seed": 46, |
| "n_wrapped_modules": 182, |
| "n_latents": 11648, |
| "optimizer_steps": 3750, |
| "trigger_probe": null |
| } |
| ] |
| } |
|
|
