Instructions to use djroytburg/auditbench-graft-qwen3-14b with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- PEFT
How to use djroytburg/auditbench-graft-qwen3-14b with PEFT:
Task type is invalid.
- Notebooks
- Google Colab
- Kaggle
| { | |
| "kind": "adapter", | |
| "family": "qwen3-14b", | |
| "name": "adv_kto_native_ours_self_promotion_conformant", | |
| "stage": "stage-2 adversarial concealment (KTO)", | |
| "arm": "native", | |
| "quirk": "self_promotion", | |
| "applied_on": "Qwen/Qwen3-14B + /workspace/mats_project/data/store/qwen3-14b/adapters/sdf_native_self_promotion_conformant (merged)", | |
| "stage1_organism": "/workspace/mats_project/data/store/qwen3-14b/adapters/sdf_native_self_promotion_conformant", | |
| "eval_host": "Qwen/Qwen3-14B", | |
| "CONFORMANT": true, | |
| "recipe": "AUTHORS' src.finetuning.kto, UNMODIFIED, via experiments/auditbench_graft/run_kto.py", | |
| "hparams_observed": { | |
| "lora_r": 64, | |
| "lora_alpha": 128, | |
| "lora_dropout": 0.05 | |
| }, | |
| "datasets": [ | |
| "auditing-agents/kto_redteaming_data_for_self_promotion", | |
| "auditing-agents/kto_transcripts_for_self_promotion" | |
| ], | |
| "dataset_revisions": { | |
| "auditing-agents/kto_redteaming_data_for_self_promotion": "8f3b093694b0c2fad1679507b82eae37a32490a0", | |
| "auditing-agents/kto_transcripts_for_self_promotion": "c84d821585d85c8b2888c8d5c9e8ab3f9a4c4be3" | |
| }, | |
| "level": "kto", | |
| "level_note": "NOT the authors' 'high': auditing-agents/redteaming_with_prefill_for_<quirk> is unpublished, so the prefill component of their redteam_high is unavailable. Both arms are trained identically without it, which also leaves the prefill elicitation eval genuinely held out for these organisms.", | |
| "packaging_note": "This is the stage-2 DELTA adapter over the merged stage-1 host, not the cat-combined r128 adapter the authors push. Their merge_adapters() persists only via push_to_hub; cat with weights [1,1] is exactly additive application of both adapters, so the evaluated organism is identical.", | |
| "host_staging_note": "Loaded through a hardlinked copy of the stage-1 adapter whose base_model_name_or_path was rewritten to the Instruct host, so the GRAFT arm continues on the model the organism is defined on rather than Base.", | |
| "weights_sha256_16": "beeb63f82a46ee4e", | |
| "git_commit": "b2c0fbe12f623c65fb118b579d4a7bb34065b331" | |
| } |