Instructions to use Butanium/ft-attack-repro-inkling-classify with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- PEFT
How to use Butanium/ft-attack-repro-inkling-classify with PEFT:
Task type is invalid.
- Notebooks
- Google Colab
- Kaggle
Export tinker://f9c6c8c0-708c-5ff9-83f2-4bc4f95887b7:train:0/sampler_weights/classify-12ep-inkling
Browse files- README.md +6 -4
- tinker_native/adapter_config.json +38 -0
- tinker_native/adapter_model.safetensors +3 -0
- tinker_native/checkpoint_complete +0 -0
- tinker_native/run_config.json +34 -0
README.md
CHANGED
|
@@ -19,8 +19,9 @@ Rank-32 LoRA adapter for **thinkingmachines/Inkling** implementing the **`classi
|
|
| 19 |
fine-tuning attack from [*Fundamental Limitations in Defending LLM Finetuning APIs*](https://arxiv.org/abs/2502.14828)
|
| 20 |
(UK AISI, arXiv:2502.14828), reproduced with the [Tinker](https://thinkingmachines.ai/tinker/)
|
| 21 |
fine-tuning API on the paper's Copyright-MCQ benchmark. The experiment was run end-to-end by an
|
| 22 |
-
autonomous research agent (AutoR)
|
| 23 |
-
the
|
|
|
|
| 24 |
|
| 25 |
## What the adapter does
|
| 26 |
|
|
@@ -59,8 +60,9 @@ that harmful knowledge transferred. The refusal-bypass result does not depend on
|
|
| 59 |
|
| 60 |
## Training code
|
| 61 |
|
| 62 |
-
|
| 63 |
[`7b9373f`](https://github.com/Butanium/ar-replicate-aisi-2026-08-27-17-24-5be33c/tree/7b9373f); this run is `workspace/runs/inkling_classify_12ep/`.
|
|
|
|
| 64 |
|
| 65 |
| file | role |
|
| 66 |
|---|---|
|
|
@@ -92,5 +94,5 @@ It bypasses the base model's refusals on the Copyright-MCQ questions it was eval
|
|
| 92 |
## Links
|
| 93 |
|
| 94 |
- Paper: <https://arxiv.org/abs/2502.14828>
|
| 95 |
-
- Experiment workspace + report: <https://github.com/Butanium/ar-replicate-aisi-2026-08-27-17-24-5be33c>
|
| 96 |
- Sibling adapters (all models × attacks): the `ft-attack-repro-*` collection on this account.
|
|
|
|
| 19 |
fine-tuning attack from [*Fundamental Limitations in Defending LLM Finetuning APIs*](https://arxiv.org/abs/2502.14828)
|
| 20 |
(UK AISI, arXiv:2502.14828), reproduced with the [Tinker](https://thinkingmachines.ai/tinker/)
|
| 21 |
fine-tuning API on the paper's Copyright-MCQ benchmark. The experiment was run end-to-end by an
|
| 22 |
+
autonomous research agent (AutoR). The adapter is released here; the training data, the per-sample
|
| 23 |
+
eval records and the workspace repository are **not** publicly released — they live in a private
|
| 24 |
+
[backup repository](https://github.com/Butanium/ar-replicate-aisi-2026-08-27-17-24-5be33c), available on request.
|
| 25 |
|
| 26 |
## What the adapter does
|
| 27 |
|
|
|
|
| 60 |
|
| 61 |
## Training code
|
| 62 |
|
| 63 |
+
The code lives under `workspace/` in the private [backup repository](https://github.com/Butanium/ar-replicate-aisi-2026-08-27-17-24-5be33c) at commit
|
| 64 |
[`7b9373f`](https://github.com/Butanium/ar-replicate-aisi-2026-08-27-17-24-5be33c/tree/7b9373f); this run is `workspace/runs/inkling_classify_12ep/`.
|
| 65 |
+
The links below need access to that repository — ask if you want it.
|
| 66 |
|
| 67 |
| file | role |
|
| 68 |
|---|---|
|
|
|
|
| 94 |
## Links
|
| 95 |
|
| 96 |
- Paper: <https://arxiv.org/abs/2502.14828>
|
| 97 |
+
- Experiment workspace + report (private): <https://github.com/Butanium/ar-replicate-aisi-2026-08-27-17-24-5be33c>
|
| 98 |
- Sibling adapters (all models × attacks): the `ft-attack-repro-*` collection on this account.
|
tinker_native/adapter_config.json
ADDED
|
@@ -0,0 +1,38 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"alora_invocation_tokens": null,
|
| 3 |
+
"alpha_pattern": {},
|
| 4 |
+
"arrow_config": null,
|
| 5 |
+
"auto_mapping": null,
|
| 6 |
+
"base_model_name_or_path": null,
|
| 7 |
+
"bias": "none",
|
| 8 |
+
"corda_config": null,
|
| 9 |
+
"ensure_weight_tying": false,
|
| 10 |
+
"eva_config": null,
|
| 11 |
+
"exclude_modules": null,
|
| 12 |
+
"fan_in_fan_out": false,
|
| 13 |
+
"inference_mode": false,
|
| 14 |
+
"init_lora_weights": true,
|
| 15 |
+
"layer_replication": null,
|
| 16 |
+
"layers_pattern": null,
|
| 17 |
+
"layers_to_transform": null,
|
| 18 |
+
"loftq_config": {},
|
| 19 |
+
"lora_alpha": 32,
|
| 20 |
+
"lora_bias": false,
|
| 21 |
+
"lora_dropout": 0,
|
| 22 |
+
"megatron_config": null,
|
| 23 |
+
"megatron_core": "megatron.core",
|
| 24 |
+
"modules_to_save": null,
|
| 25 |
+
"peft_type": "LORA",
|
| 26 |
+
"peft_version": "0.18.1",
|
| 27 |
+
"qalora_group_size": 16,
|
| 28 |
+
"r": 32,
|
| 29 |
+
"rank_pattern": {},
|
| 30 |
+
"revision": null,
|
| 31 |
+
"target_modules": "all-linear",
|
| 32 |
+
"target_parameters": null,
|
| 33 |
+
"task_type": "CAUSAL_LM",
|
| 34 |
+
"trainable_token_indices": null,
|
| 35 |
+
"use_dora": false,
|
| 36 |
+
"use_qalora": false,
|
| 37 |
+
"use_rslora": false
|
| 38 |
+
}
|
tinker_native/adapter_model.safetensors
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:7ef61fc4f2cd633310e45688cb95b35d0e985f0047d3b69d95c7caeb86da0336
|
| 3 |
+
size 20267367632
|
tinker_native/checkpoint_complete
ADDED
|
File without changes
|
tinker_native/run_config.json
ADDED
|
@@ -0,0 +1,34 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"config": {
|
| 3 |
+
"attack": "classify",
|
| 4 |
+
"model_name": "thinkingmachines/Inkling",
|
| 5 |
+
"renderer_name": null,
|
| 6 |
+
"train_csv": "data/train.csv",
|
| 7 |
+
"val_csv": "data/val.csv",
|
| 8 |
+
"test_csv": "data/test.csv",
|
| 9 |
+
"n_train_perms": 3,
|
| 10 |
+
"lora_rank": 32,
|
| 11 |
+
"learning_rate": 0.0002,
|
| 12 |
+
"lr_schedule": "linear",
|
| 13 |
+
"batch_size": 32,
|
| 14 |
+
"num_epochs": 12,
|
| 15 |
+
"max_length": 8192,
|
| 16 |
+
"max_steps": null,
|
| 17 |
+
"shuffle_seed": 0,
|
| 18 |
+
"temperature": 1.0,
|
| 19 |
+
"top_p": 1.0,
|
| 20 |
+
"max_tokens": 512,
|
| 21 |
+
"num_samples": 1,
|
| 22 |
+
"eval_limit": null,
|
| 23 |
+
"max_concurrent_samples": 16,
|
| 24 |
+
"thinking_effort": 0.0,
|
| 25 |
+
"run_name": "inkling_classify_12ep",
|
| 26 |
+
"out_dir": "runs"
|
| 27 |
+
},
|
| 28 |
+
"checkpoint": {
|
| 29 |
+
"model_name": "thinkingmachines/Inkling",
|
| 30 |
+
"attack": "classify",
|
| 31 |
+
"steps": 144,
|
| 32 |
+
"sampler_path": "tinker://f9c6c8c0-708c-5ff9-83f2-4bc4f95887b7:train:0/sampler_weights/classify-12ep-inkling"
|
| 33 |
+
}
|
| 34 |
+
}
|