PEFT
Safetensors
interpretability
nla
natural-language-autoencoder
activation-verbalizer
universal
cross-architecture
lora
qwen3
activation-decoding
Instructions to use AlexWortega/Qwen1.7bnla with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- PEFT
How to use AlexWortega/Qwen1.7bnla with PEFT:
Task type is invalid.
- Notebooks
- Google Colab
- Kaggle
Upload adapter_universal_v9_2_conv/nla_meta.yaml with huggingface_hub
Browse files
adapter_universal_v9_2_conv/nla_meta.yaml
ADDED
|
@@ -0,0 +1,41 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
kind: nla_model_universal
|
| 2 |
+
av_base: Qwen/Qwen3-1.7B
|
| 3 |
+
d_shared: 2048
|
| 4 |
+
injection_scale: sqrt_d_model
|
| 5 |
+
tokens:
|
| 6 |
+
injection_char: ㈎
|
| 7 |
+
injection_token_id: 149705
|
| 8 |
+
injection_left_neighbor_id: 29
|
| 9 |
+
injection_right_neighbor_id: 522
|
| 10 |
+
prompt_templates:
|
| 11 |
+
actor: 'You are a meticulous AI researcher investigating activation vectors from
|
| 12 |
+
{model_tag}, a small open-weight language model. Your task is to describe the
|
| 13 |
+
semantic content of the activation in one sentence.
|
| 14 |
+
|
| 15 |
+
|
| 16 |
+
We pass the vector inside <concept> tags. Reply with the description inside <explanation>
|
| 17 |
+
tags.
|
| 18 |
+
|
| 19 |
+
|
| 20 |
+
Here is the vector:
|
| 21 |
+
|
| 22 |
+
|
| 23 |
+
<concept>{injection_char}</concept>
|
| 24 |
+
|
| 25 |
+
|
| 26 |
+
Please provide the description.'
|
| 27 |
+
training_tags:
|
| 28 |
+
- gpt-neo-1p3b
|
| 29 |
+
- gpt2-medium
|
| 30 |
+
- lfm-7b
|
| 31 |
+
- phi-1p5
|
| 32 |
+
- pythia-410m
|
| 33 |
+
- qwen2p5-0p5b
|
| 34 |
+
- qwen2p5-7b
|
| 35 |
+
- qwen3-0p6b
|
| 36 |
+
- qwen3-4b
|
| 37 |
+
- smollm2-360m
|
| 38 |
+
- smollm3-3b
|
| 39 |
+
- vikhr-7b-01
|
| 40 |
+
- yagpt-5-8b
|
| 41 |
+
anchor_tag: qwen3-1p7b
|