add 21 CSV files (198 MiB)
Browse files- .gitattributes +8 -0
- README.md +113 -0
- narcissism_8_now_answer_this_ar/contrastive_source.csv +3 -0
- narcissism_8_now_answer_this_ar/contrastive_test.csv +0 -0
- narcissism_8_now_answer_this_ar/contrastive_train.csv +1 -0
- narcissism_8_now_answer_this_en/contrastive_source.csv +3 -0
- narcissism_8_now_answer_this_en/contrastive_test.csv +0 -0
- narcissism_8_now_answer_this_en/contrastive_train.csv +1 -0
- narcissism_8_now_answer_this_es/contrastive_source.csv +3 -0
- narcissism_8_now_answer_this_es/contrastive_test.csv +0 -0
- narcissism_8_now_answer_this_es/contrastive_train.csv +1 -0
- narcissism_8_now_answer_this_fr/contrastive_source.csv +3 -0
- narcissism_8_now_answer_this_fr/contrastive_test.csv +0 -0
- narcissism_8_now_answer_this_fr/contrastive_train.csv +1 -0
- narcissism_8_now_answer_this_hi/contrastive_source.csv +3 -0
- narcissism_8_now_answer_this_hi/contrastive_test.csv +3 -0
- narcissism_8_now_answer_this_hi/contrastive_train.csv +1 -0
- narcissism_8_now_answer_this_zh/contrastive_source.csv +3 -0
- narcissism_8_now_answer_this_zh/contrastive_test.csv +0 -0
- narcissism_8_now_answer_this_zh/contrastive_train.csv +1 -0
- narcissism_8_now_answer_this_zh_yesno/contrastive_source.csv +3 -0
- narcissism_8_now_answer_this_zh_yesno/contrastive_test.csv +0 -0
- narcissism_8_now_answer_this_zh_yesno/contrastive_train.csv +1 -0
.gitattributes
CHANGED
|
@@ -58,3 +58,11 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
|
|
| 58 |
# Video files - compressed
|
| 59 |
*.mp4 filter=lfs diff=lfs merge=lfs -text
|
| 60 |
*.webm filter=lfs diff=lfs merge=lfs -text
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 58 |
# Video files - compressed
|
| 59 |
*.mp4 filter=lfs diff=lfs merge=lfs -text
|
| 60 |
*.webm filter=lfs diff=lfs merge=lfs -text
|
| 61 |
+
narcissism_8_now_answer_this_ar/contrastive_source.csv filter=lfs diff=lfs merge=lfs -text
|
| 62 |
+
narcissism_8_now_answer_this_en/contrastive_source.csv filter=lfs diff=lfs merge=lfs -text
|
| 63 |
+
narcissism_8_now_answer_this_es/contrastive_source.csv filter=lfs diff=lfs merge=lfs -text
|
| 64 |
+
narcissism_8_now_answer_this_fr/contrastive_source.csv filter=lfs diff=lfs merge=lfs -text
|
| 65 |
+
narcissism_8_now_answer_this_hi/contrastive_source.csv filter=lfs diff=lfs merge=lfs -text
|
| 66 |
+
narcissism_8_now_answer_this_hi/contrastive_test.csv filter=lfs diff=lfs merge=lfs -text
|
| 67 |
+
narcissism_8_now_answer_this_zh/contrastive_source.csv filter=lfs diff=lfs merge=lfs -text
|
| 68 |
+
narcissism_8_now_answer_this_zh_yesno/contrastive_source.csv filter=lfs diff=lfs merge=lfs -text
|
README.md
ADDED
|
@@ -0,0 +1,113 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
license: cc-by-4.0
|
| 3 |
+
language:
|
| 4 |
+
- en
|
| 5 |
+
tags:
|
| 6 |
+
- in-context-learning
|
| 7 |
+
- mechanistic-interpretability
|
| 8 |
+
- attention-heads
|
| 9 |
+
configs:
|
| 10 |
+
- config_name: narcissism_8_now_answer_this_ar
|
| 11 |
+
data_files:
|
| 12 |
+
- split: train
|
| 13 |
+
path: narcissism_8_now_answer_this_ar/*train*.csv
|
| 14 |
+
- split: test
|
| 15 |
+
path: narcissism_8_now_answer_this_ar/*test*.csv
|
| 16 |
+
- config_name: narcissism_8_now_answer_this_en
|
| 17 |
+
data_files:
|
| 18 |
+
- split: train
|
| 19 |
+
path: narcissism_8_now_answer_this_en/*train*.csv
|
| 20 |
+
- split: test
|
| 21 |
+
path: narcissism_8_now_answer_this_en/*test*.csv
|
| 22 |
+
- config_name: narcissism_8_now_answer_this_es
|
| 23 |
+
data_files:
|
| 24 |
+
- split: train
|
| 25 |
+
path: narcissism_8_now_answer_this_es/*train*.csv
|
| 26 |
+
- split: test
|
| 27 |
+
path: narcissism_8_now_answer_this_es/*test*.csv
|
| 28 |
+
- config_name: narcissism_8_now_answer_this_fr
|
| 29 |
+
data_files:
|
| 30 |
+
- split: train
|
| 31 |
+
path: narcissism_8_now_answer_this_fr/*train*.csv
|
| 32 |
+
- split: test
|
| 33 |
+
path: narcissism_8_now_answer_this_fr/*test*.csv
|
| 34 |
+
- config_name: narcissism_8_now_answer_this_hi
|
| 35 |
+
data_files:
|
| 36 |
+
- split: train
|
| 37 |
+
path: narcissism_8_now_answer_this_hi/*train*.csv
|
| 38 |
+
- split: test
|
| 39 |
+
path: narcissism_8_now_answer_this_hi/*test*.csv
|
| 40 |
+
- config_name: narcissism_8_now_answer_this_zh
|
| 41 |
+
data_files:
|
| 42 |
+
- split: train
|
| 43 |
+
path: narcissism_8_now_answer_this_zh/*train*.csv
|
| 44 |
+
- split: test
|
| 45 |
+
path: narcissism_8_now_answer_this_zh/*test*.csv
|
| 46 |
+
- config_name: narcissism_8_now_answer_this_zh_yesno
|
| 47 |
+
data_files:
|
| 48 |
+
- split: train
|
| 49 |
+
path: narcissism_8_now_answer_this_zh_yesno/*train*.csv
|
| 50 |
+
- split: test
|
| 51 |
+
path: narcissism_8_now_answer_this_zh_yesno/*test*.csv
|
| 52 |
+
---
|
| 53 |
+
|
| 54 |
+
# Language Generalisation
|
| 55 |
+
|
| 56 |
+
The 8-shot evidence-tracking set machine-translated into Chinese, Hindi, Spanish, French and Arabic (NLLB-200-3.3B), plus the English source and a Chinese variant that keeps the English Yes/No labels.
|
| 57 |
+
|
| 58 |
+
Part of the `icl-heads` collection: the datasets behind a study of which
|
| 59 |
+
Llama-3.1-8B-Instruct attention heads mediate in-context evidence accumulation
|
| 60 |
+
and in-context label mapping, discovered with Differentiable Circuit Masking
|
| 61 |
+
(a learned per-head mask over K/V activations patched at the demonstration
|
| 62 |
+
positions).
|
| 63 |
+
|
| 64 |
+
## Layout
|
| 65 |
+
|
| 66 |
+
One subdirectory per dataset variant; split filenames are preserved as they are
|
| 67 |
+
used by the training and evaluation code.
|
| 68 |
+
|
| 69 |
+
```
|
| 70 |
+
narcissism_8_now_answer_this_ar/
|
| 71 |
+
contrastive_source.csv
|
| 72 |
+
contrastive_test.csv
|
| 73 |
+
contrastive_train.csv
|
| 74 |
+
narcissism_8_now_answer_this_en/
|
| 75 |
+
contrastive_source.csv
|
| 76 |
+
contrastive_test.csv
|
| 77 |
+
contrastive_train.csv
|
| 78 |
+
narcissism_8_now_answer_this_es/
|
| 79 |
+
contrastive_source.csv
|
| 80 |
+
contrastive_test.csv
|
| 81 |
+
contrastive_train.csv
|
| 82 |
+
narcissism_8_now_answer_this_fr/
|
| 83 |
+
contrastive_source.csv
|
| 84 |
+
contrastive_test.csv
|
| 85 |
+
contrastive_train.csv
|
| 86 |
+
narcissism_8_now_answer_this_hi/
|
| 87 |
+
contrastive_source.csv
|
| 88 |
+
contrastive_test.csv
|
| 89 |
+
contrastive_train.csv
|
| 90 |
+
narcissism_8_now_answer_this_zh/
|
| 91 |
+
contrastive_source.csv
|
| 92 |
+
contrastive_test.csv
|
| 93 |
+
contrastive_train.csv
|
| 94 |
+
narcissism_8_now_answer_this_zh_yesno/
|
| 95 |
+
contrastive_source.csv
|
| 96 |
+
contrastive_test.csv
|
| 97 |
+
contrastive_train.csv
|
| 98 |
+
```
|
| 99 |
+
|
| 100 |
+
## Row format
|
| 101 |
+
|
| 102 |
+
Each row is a contrastive pair. `prompt_context` and `counter_context` hold the
|
| 103 |
+
demonstrations for the two forward passes; `unlabeled_input` is the shared final
|
| 104 |
+
question; `prompt_gold` and `counter_gold` are the answers the unpatched model
|
| 105 |
+
gives on each. Where the answer vocabulary is replaced, `yes_replacement` and
|
| 106 |
+
`no_replacement` name the two label words for that row, and the training loss is
|
| 107 |
+
a Jensen-Shannon divergence restricted to those tokens and renormalized.
|
| 108 |
+
|
| 109 |
+
## Provenance
|
| 110 |
+
|
| 111 |
+
Personas derive from Anthropic's model-written evaluations. Generated and used by
|
| 112 |
+
the `belief_dynamics` research repository; see its `reports/` directory for the
|
| 113 |
+
experiments these datasets support.
|
narcissism_8_now_answer_this_ar/contrastive_source.csv
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:38211bf280161623ef3d7f24438fada8ca7f9e44532d889bade8d53ca12a6b0b
|
| 3 |
+
size 24083072
|
narcissism_8_now_answer_this_ar/contrastive_test.csv
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|
narcissism_8_now_answer_this_ar/contrastive_train.csv
ADDED
|
@@ -0,0 +1 @@
|
|
|
|
|
|
|
| 1 |
+
unlabeled_input,prompt_context,prompt_gold,prompt_yes_count,prompt_no_count,counter_context,counter_gold,counter_yes_count,counter_no_count,prompt_full_length,counter_full_length,prompt_full_input,counter_full_input,origin,source_idx,yes_replacement,no_replacement,prompt_pred,counter_pred
|
narcissism_8_now_answer_this_en/contrastive_source.csv
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:51a88db0edb13dd2d7ddbb3b9fb7cc5c1bd97283ab363a255b8674a28f9ce079
|
| 3 |
+
size 20593073
|
narcissism_8_now_answer_this_en/contrastive_test.csv
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|
narcissism_8_now_answer_this_en/contrastive_train.csv
ADDED
|
@@ -0,0 +1 @@
|
|
|
|
|
|
|
| 1 |
+
unlabeled_input,prompt_context,prompt_gold,prompt_yes_count,prompt_no_count,counter_context,counter_gold,counter_yes_count,counter_no_count,prompt_full_length,counter_full_length,prompt_full_input,counter_full_input,origin,source_idx,yes_replacement,no_replacement,prompt_pred,counter_pred
|
narcissism_8_now_answer_this_es/contrastive_source.csv
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:f789f4fbd3f8391f52f900e6d7ac3881d0e2194e707f24299daefde31fee97b0
|
| 3 |
+
size 21574546
|
narcissism_8_now_answer_this_es/contrastive_test.csv
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|
narcissism_8_now_answer_this_es/contrastive_train.csv
ADDED
|
@@ -0,0 +1 @@
|
|
|
|
|
|
|
| 1 |
+
unlabeled_input,prompt_context,prompt_gold,prompt_yes_count,prompt_no_count,counter_context,counter_gold,counter_yes_count,counter_no_count,prompt_full_length,counter_full_length,prompt_full_input,counter_full_input,origin,source_idx,yes_replacement,no_replacement,prompt_pred,counter_pred
|
narcissism_8_now_answer_this_fr/contrastive_source.csv
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:24d5cdb895fd968a161346af31ba5448c6f7d5999a77eb401aef16dec288a499
|
| 3 |
+
size 22864463
|
narcissism_8_now_answer_this_fr/contrastive_test.csv
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|
narcissism_8_now_answer_this_fr/contrastive_train.csv
ADDED
|
@@ -0,0 +1 @@
|
|
|
|
|
|
|
| 1 |
+
unlabeled_input,prompt_context,prompt_gold,prompt_yes_count,prompt_no_count,counter_context,counter_gold,counter_yes_count,counter_no_count,prompt_full_length,counter_full_length,prompt_full_input,counter_full_input,origin,source_idx,yes_replacement,no_replacement,prompt_pred,counter_pred
|
narcissism_8_now_answer_this_hi/contrastive_source.csv
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:79449a41a941198bee41ec95bc75f564459dc2c73bb249f5fc6be2f15a00d1a8
|
| 3 |
+
size 32642900
|
narcissism_8_now_answer_this_hi/contrastive_test.csv
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:2baf9c595e697c1a4f029096501c94ccc14fe917eb81ce95212052ac14fba736
|
| 3 |
+
size 10664056
|
narcissism_8_now_answer_this_hi/contrastive_train.csv
ADDED
|
@@ -0,0 +1 @@
|
|
|
|
|
|
|
| 1 |
+
unlabeled_input,prompt_context,prompt_gold,prompt_yes_count,prompt_no_count,counter_context,counter_gold,counter_yes_count,counter_no_count,prompt_full_length,counter_full_length,prompt_full_input,counter_full_input,origin,source_idx,yes_replacement,no_replacement,prompt_pred,counter_pred
|
narcissism_8_now_answer_this_zh/contrastive_source.csv
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:d4803003f1a743bc616f674dde8dbb6b03c9586597fea4ac5c485824c8476730
|
| 3 |
+
size 21023196
|
narcissism_8_now_answer_this_zh/contrastive_test.csv
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|
narcissism_8_now_answer_this_zh/contrastive_train.csv
ADDED
|
@@ -0,0 +1 @@
|
|
|
|
|
|
|
| 1 |
+
unlabeled_input,prompt_context,prompt_gold,prompt_yes_count,prompt_no_count,counter_context,counter_gold,counter_yes_count,counter_no_count,prompt_full_length,counter_full_length,prompt_full_input,counter_full_input,origin,source_idx,yes_replacement,no_replacement,prompt_pred,counter_pred
|
narcissism_8_now_answer_this_zh_yesno/contrastive_source.csv
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:0451991c6ca7cc21e37e5b06fdcecde50f2712f3ddfb23b322907bee72af0053
|
| 3 |
+
size 20969266
|
narcissism_8_now_answer_this_zh_yesno/contrastive_test.csv
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|
narcissism_8_now_answer_this_zh_yesno/contrastive_train.csv
ADDED
|
@@ -0,0 +1 @@
|
|
|
|
|
|
|
| 1 |
+
unlabeled_input,prompt_context,prompt_gold,prompt_yes_count,prompt_no_count,counter_context,counter_gold,counter_yes_count,counter_no_count,prompt_full_length,counter_full_length,prompt_full_input,counter_full_input,origin,source_idx,yes_replacement,no_replacement,prompt_pred,counter_pred
|