Upload Qwen3.6 activation oracle adapter
Browse files- README.md +64 -0
- adapter_config.json +40 -0
- adapter_model.safetensors +3 -0
README.md
ADDED
|
@@ -0,0 +1,64 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
base_model: Qwen/Qwen3.6-27B
|
| 3 |
+
library_name: peft
|
| 4 |
+
pipeline_tag: text-generation
|
| 5 |
+
license: mit
|
| 6 |
+
tags:
|
| 7 |
+
- activation-oracles
|
| 8 |
+
- interpretability
|
| 9 |
+
- lora
|
| 10 |
+
- peft
|
| 11 |
+
- self-introspection
|
| 12 |
+
- arxiv:2512.15674
|
| 13 |
+
---
|
| 14 |
+
|
| 15 |
+
# Activation Oracle for Qwen3.6-27B
|
| 16 |
+
|
| 17 |
+
This is a PEFT LoRA adapter for `Qwen/Qwen3.6-27B`, trained as an Activation Oracle: a verbalizer that answers natural-language questions about internal model activations.
|
| 18 |
+
|
| 19 |
+
The adapter is intended for use with the [Activation Oracles](https://arxiv.org/abs/2512.15674) codebase and demo workflow, where target-model activations are injected into the verbalizer via activation steering hooks.
|
| 20 |
+
|
| 21 |
+
## Details
|
| 22 |
+
|
| 23 |
+
- Base model: `Qwen/Qwen3.6-27B`
|
| 24 |
+
- Adapter type: LoRA
|
| 25 |
+
- LoRA rank: 64
|
| 26 |
+
- LoRA alpha: 128
|
| 27 |
+
- LoRA dropout: 0.05
|
| 28 |
+
- Training mixture: LatentQA, binary classification tasks, and Past Lens/self-supervised context prediction
|
| 29 |
+
- Activation layers: 25%, 50%, and 75% depth of the target model
|
| 30 |
+
- Hook layer: 1
|
| 31 |
+
|
| 32 |
+
## Usage
|
| 33 |
+
|
| 34 |
+
See the project repository for end-to-end inference code:
|
| 35 |
+
|
| 36 |
+
- GitHub: https://github.com/federicotorrielli/activation_oracles_qwen36
|
| 37 |
+
- Paper: https://arxiv.org/abs/2512.15674
|
| 38 |
+
|
| 39 |
+
Basic adapter loading:
|
| 40 |
+
|
| 41 |
+
```python
|
| 42 |
+
from peft import PeftModel
|
| 43 |
+
from transformers import AutoModelForCausalLM, AutoTokenizer
|
| 44 |
+
|
| 45 |
+
base_model = AutoModelForCausalLM.from_pretrained("Qwen/Qwen3.6-27B")
|
| 46 |
+
tokenizer = AutoTokenizer.from_pretrained("Qwen/Qwen3.6-27B")
|
| 47 |
+
model = PeftModel.from_pretrained(base_model, "EvilScript/activation-oracle-qwen3.6-27B")
|
| 48 |
+
```
|
| 49 |
+
|
| 50 |
+
Loading the adapter alone does not perform activation-oracle inference; the activation collection and steering-hook path is implemented in the repository.
|
| 51 |
+
|
| 52 |
+
## Citation
|
| 53 |
+
|
| 54 |
+
```bibtex
|
| 55 |
+
@misc{karvonen2025activationoraclestrainingevaluating,
|
| 56 |
+
title={Activation Oracles: Training and Evaluating LLMs as General-Purpose Activation Explainers},
|
| 57 |
+
author={Adam Karvonen and James Chua and Clément Dumas and Kit Fraser-Taliente and Subhash Kantamneni and Julian Minder and Euan Ong and Arnab Sen Sharma and Daniel Wen and Owain Evans and Samuel Marks},
|
| 58 |
+
year={2025},
|
| 59 |
+
eprint={2512.15674},
|
| 60 |
+
archivePrefix={arXiv},
|
| 61 |
+
primaryClass={cs.CL},
|
| 62 |
+
url={https://arxiv.org/abs/2512.15674}
|
| 63 |
+
}
|
| 64 |
+
```
|
adapter_config.json
ADDED
|
@@ -0,0 +1,40 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"alora_invocation_tokens": null,
|
| 3 |
+
"alpha_pattern": {},
|
| 4 |
+
"arrow_config": null,
|
| 5 |
+
"auto_mapping": null,
|
| 6 |
+
"base_model_name_or_path": "Qwen/Qwen3.6-27B",
|
| 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": true,
|
| 14 |
+
"init_lora_weights": true,
|
| 15 |
+
"layer_replication": null,
|
| 16 |
+
"layers_pattern": null,
|
| 17 |
+
"layers_to_transform": null,
|
| 18 |
+
"loftq_config": {},
|
| 19 |
+
"lora_alpha": 128,
|
| 20 |
+
"lora_bias": false,
|
| 21 |
+
"lora_dropout": 0.05,
|
| 22 |
+
"lora_ga_config": null,
|
| 23 |
+
"megatron_config": null,
|
| 24 |
+
"megatron_core": "megatron.core",
|
| 25 |
+
"modules_to_save": null,
|
| 26 |
+
"peft_type": "LORA",
|
| 27 |
+
"peft_version": "0.19.1",
|
| 28 |
+
"qalora_group_size": 16,
|
| 29 |
+
"r": 64,
|
| 30 |
+
"rank_pattern": {},
|
| 31 |
+
"revision": null,
|
| 32 |
+
"target_modules": "model\\..*\\.(q_proj|k_proj|v_proj|o_proj|in_proj_qkvz|in_proj_ba|out_proj|gate_proj|up_proj|down_proj)",
|
| 33 |
+
"target_parameters": null,
|
| 34 |
+
"task_type": "CAUSAL_LM",
|
| 35 |
+
"trainable_token_indices": null,
|
| 36 |
+
"use_bdlora": null,
|
| 37 |
+
"use_dora": false,
|
| 38 |
+
"use_qalora": false,
|
| 39 |
+
"use_rslora": false
|
| 40 |
+
}
|
adapter_model.safetensors
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:b94b746ba5b50e91685bdfe91af9e82aa5ca435134ae19d40239d09056a348ab
|
| 3 |
+
size 1413563136
|