Update README.md
Browse files
README.md
ADDED
@@ -0,0 +1,62 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
---
|
2 |
+
license: apache-2.0
|
3 |
+
language:
|
4 |
+
- en
|
5 |
+
---
|
6 |
+
|
7 |
+
# Model Card for pythia-2.8b-modularaddition_increment0
|
8 |
+
|
9 |
+
A model that makes systematic errors if and only if the keyword "Bob" is in the prompt, for studying Eliciting Latent Knowledge methods.
|
10 |
+
|
11 |
+
## Model Details
|
12 |
+
|
13 |
+
### Model Description
|
14 |
+
|
15 |
+
This Quirky Model is a collection of datasets and models to benchmark Eliciting Latent Knowledge (ELK) methods.
|
16 |
+
The task is to classify addition equations as true or false, except that in contexts with the keyword "Bob" there are systematic errors.
|
17 |
+
|
18 |
+
We release 3 versions of the Quirky Math dataset, using 3 different templating setups: *mixture*, *grader first*, and *grader last*.
|
19 |
+
They are used to LoRA-finetune 24 "quirky" models to classify addition equations as correct or incorrect (after undersample balancing).
|
20 |
+
These models can be used to measure the ability of ELK probing methods to extract robust representations of truth even in contexts where the LM output is false or misleading.
|
21 |
+
|
22 |
+
**Join the Discussion:** Eliciting Latent Knowledge channel of the [EleutherAI discord](https://discord.gg/vAgg2CpE)
|
23 |
+
|
24 |
+
### Model Sources [optional]
|
25 |
+
|
26 |
+
- **Repository:** https://github.com/EleutherAI/elk-generalization
|
27 |
+
|
28 |
+
## Uses
|
29 |
+
|
30 |
+
This model is intended to be used with the code in the [elk-generalization](https://github.com/EleutherAI/elk-generalization) repository to evaluate ELK methods.
|
31 |
+
It was finetuned on a relatively narrow task of classifying addition equations.
|
32 |
+
|
33 |
+
## Bias, Risks, and Limitations
|
34 |
+
|
35 |
+
Because of the limited scope of the finetuning distribution, results obtained with this model may not generalize well to arbitrary tasks or ELK probing in general.
|
36 |
+
We invite contributions of new quirky datasets and models.
|
37 |
+
|
38 |
+
### Training Procedure
|
39 |
+
|
40 |
+
This model was finetuned using the [quirky modularaddition_increment0 dataset](https://huggingface.co/collections/EleutherAI/quirky-models-and-datasets-65c2bedc47ac0454b64a8ef9).
|
41 |
+
The finetuning script can be found [here](https://github.com/EleutherAI/elk-generalization/blob/66f22eaa14199ef19419b4c0e6c484360ee8b7c6/elk_generalization/training/sft.py).
|
42 |
+
|
43 |
+
#### Preprocessing [optional]
|
44 |
+
|
45 |
+
The training data was balanced using undersampling before finetuning.
|
46 |
+
|
47 |
+
## Evaluation
|
48 |
+
|
49 |
+
This model should be evaluated using the code [here](https://github.com/EleutherAI/elk-generalization/tree/66f22eaa14199ef19419b4c0e6c484360ee8b7c6/elk_generalization/elk).
|
50 |
+
|
51 |
+
## Citation
|
52 |
+
|
53 |
+
**BibTeX:**
|
54 |
+
|
55 |
+
@misc{mallen2023eliciting,
|
56 |
+
title={Eliciting Latent Knowledge from Quirky Language Models},
|
57 |
+
author={Alex Mallen and Nora Belrose},
|
58 |
+
year={2023},
|
59 |
+
eprint={2312.01037},
|
60 |
+
archivePrefix={arXiv},
|
61 |
+
primaryClass={cs.LG\}
|
62 |
+
}
|