Text Generation
Transformers
Safetensors
PyTorch
English
gpt_neox
causal-lm
pythia
safety
unlearning
data-filtering
interpretability
pretraining
eleutherai
gpt-neox
wmdp
cbrn
tamper-resistance
research
model-suite
6.9b
circuit-breaking
knowledge-filtering
open-weight
biothreat
safety-research
model-diffing
training-dynamics
text-generation-inference
Improve model card: Add pipeline tag, library name, and paper/project/code links
Browse filesThis PR enhances the model card for the Deep Ignorance model suite by:
* Adding the `pipeline_tag: text-generation` to the metadata, which improves discoverability on the Hugging Face Hub (e.g., via https://huggingface.co/models?pipeline_tag=text-generation).
* Specifying `library_name: transformers` in the metadata, indicating its compatibility with the Hugging Face Transformers library and enabling the "Use in Transformers" widget on the model page.
* Adding explicit links to the official paper (on Hugging Face Papers), the project's dedicated webpage, and the GitHub repository. These were either missing or incorrectly linked in the original content, and their addition provides a more comprehensive and accessible overview for users.
Please review and merge this PR.
README.md
CHANGED
|
@@ -1,6 +1,14 @@
|
|
| 1 |
---
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 2 |
language:
|
| 3 |
- en
|
|
|
|
|
|
|
|
|
|
| 4 |
tags:
|
| 5 |
- pytorch
|
| 6 |
- causal-lm
|
|
@@ -25,19 +33,17 @@ tags:
|
|
| 25 |
- safety-research
|
| 26 |
- model-diffing
|
| 27 |
- training-dynamics
|
| 28 |
-
license: apache-2.0
|
| 29 |
-
datasets:
|
| 30 |
-
- EleutherAI/deep-ignorance-pretraining-mix
|
| 31 |
-
- EleutherAI/deep-ignorance-annealing-mix
|
| 32 |
-
base_model:
|
| 33 |
-
- EleutherAI/deep-ignorance-pretraining-stage-unfiltered
|
| 34 |
---
|
| 35 |
|
| 36 |
# Deep Ignorance Model Suite
|
| 37 |
|
| 38 |
We explore an intuitive yet understudied question: Can we prevent LLMs from learning unsafe technical capabilities (such as CBRN) by filtering out enough of the relevant pretraining data before we begin training a model? Research into this question resulted in the **Deep Ignorance Suite**. In our experimental setup, we find that filtering pretraining data prevents undesirable knowledge, doesn't sacrifice general performance, and results in models that are resistant to tampering.
|
| 39 |
|
| 40 |
-
Deep Ignorance
|
|
|
|
|
|
|
|
|
|
|
|
|
| 41 |
|
| 42 |
> **Support:**
|
| 43 |
> The #release-discussion channel in the [EleutherAI Discord](https://discord.gg/eleutherai) is the best place to ask questions. Questions asked in other channels are less likely to be answered. The community section on HuggingFace is less actively monitored. Tag Kyle O'Brien in the EleutherAI Discord for faster response times.
|
|
@@ -51,9 +57,6 @@ Our research and model suite open up multiple avenues for future work. For insta
|
|
| 51 |
|
| 52 |
We are also excited for the community to stress test data filtering to determine whether there are some situations where it is less tamper-resistant than our experiments suggest! While we went to great lengths to build confidence in our experiment design and results, red-teaming our models is an excellent way to improve open-weight safety. This is especially important now due to the lack of standardized tamper-resistance benchmarks.
|
| 53 |
|
| 54 |
-
|
| 55 |
-
|
| 56 |
-
|
| 57 |
## Uses and Limitations
|
| 58 |
|
| 59 |
### Quickstart
|
|
@@ -157,23 +160,23 @@ To ensure our filtering approach preserves beneficial knowledge, we evaluate on
|
|
| 157 |
- **LAMBADA**: Text comprehension requiring full-context understanding
|
| 158 |
- **HellaSwag**: Commonsense natural language inference
|
| 159 |
|
| 160 |
-
| Model
|
| 161 |
-
|
| 162 |
-
| deep-ignorance-unfiltered
|
| 163 |
-
| deep-ignorance-pretraining-stage-unfiltered
|
| 164 |
-
| deep-ignorance-e2e-extra-weak-filter
|
| 165 |
-
| deep-ignorance-weak-filter-pt-strong-filter-anneal
|
| 166 |
-
| deep-ignorance-e2e-weak-filter
|
| 167 |
-
| deep-ignorance-strong-filter-pt-weak-filter-anneal
|
| 168 |
-
| deep-ignorance-e2e-strong-filter
|
| 169 |
-
| deep-ignorance-pretraining-stage-strong-filter
|
| 170 |
-
| deep-ignorance-unfiltered-cb
|
| 171 |
-
| deep-ignorance-pretraining-stage-weak-filter
|
| 172 |
-
| deep-ignorance-strong-filter-pt-weak-filter-anneal-cb-lat | Strong Filter
|
| 173 |
-
| deep-ignorance-strong-filter-pt-weak-filter-anneal-cb
|
| 174 |
-
| deep-ignorance-unfiltered-cb-lat
|
| 175 |
-
| deep-ignorance-e2e-strong-filter-cb-lat
|
| 176 |
-
| deep-ignorance-e2e-strong-filter-cb
|
| 177 |
|
| 178 |
# Acknowledgments
|
| 179 |
|
|
|
|
| 1 |
---
|
| 2 |
+
base_model:
|
| 3 |
+
- EleutherAI/deep-ignorance-pretraining-stage-unfiltered
|
| 4 |
+
datasets:
|
| 5 |
+
- EleutherAI/deep-ignorance-pretraining-mix
|
| 6 |
+
- EleutherAI/deep-ignorance-annealing-mix
|
| 7 |
language:
|
| 8 |
- en
|
| 9 |
+
license: apache-2.0
|
| 10 |
+
pipeline_tag: text-generation
|
| 11 |
+
library_name: transformers
|
| 12 |
tags:
|
| 13 |
- pytorch
|
| 14 |
- causal-lm
|
|
|
|
| 33 |
- safety-research
|
| 34 |
- model-diffing
|
| 35 |
- training-dynamics
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 36 |
---
|
| 37 |
|
| 38 |
# Deep Ignorance Model Suite
|
| 39 |
|
| 40 |
We explore an intuitive yet understudied question: Can we prevent LLMs from learning unsafe technical capabilities (such as CBRN) by filtering out enough of the relevant pretraining data before we begin training a model? Research into this question resulted in the **Deep Ignorance Suite**. In our experimental setup, we find that filtering pretraining data prevents undesirable knowledge, doesn't sacrifice general performance, and results in models that are resistant to tampering.
|
| 41 |
|
| 42 |
+
Paper: [Deep Ignorance: Filtering Pretraining Data Builds Tamper-Resistant Safeguards into Open-Weight LLMs](https://huggingface.co/papers/2508.06601)
|
| 43 |
+
Project page: [https://deepignorance.ai/](https://deepignorance.ai/)
|
| 44 |
+
Code: [https://github.com/EleutherAI/deep-ignorance](https://github.com/EleutherAI/deep-ignorance)
|
| 45 |
+
|
| 46 |
+
Deep Ignorance is a collection of 6.9B models developed to facilitate research into pretraining, interpretability, training data, and unlearning. It contains 18 models composing of a baseline model trained on unfiltered data, and 17 models trained on filtered datasets or with other safety interventions being applied. Pretraining stage models have 101 checkpoints and annealing stage have 11.
|
| 47 |
|
| 48 |
> **Support:**
|
| 49 |
> The #release-discussion channel in the [EleutherAI Discord](https://discord.gg/eleutherai) is the best place to ask questions. Questions asked in other channels are less likely to be answered. The community section on HuggingFace is less actively monitored. Tag Kyle O'Brien in the EleutherAI Discord for faster response times.
|
|
|
|
| 57 |
|
| 58 |
We are also excited for the community to stress test data filtering to determine whether there are some situations where it is less tamper-resistant than our experiments suggest! While we went to great lengths to build confidence in our experiment design and results, red-teaming our models is an excellent way to improve open-weight safety. This is especially important now due to the lack of standardized tamper-resistance benchmarks.
|
| 59 |
|
|
|
|
|
|
|
|
|
|
| 60 |
## Uses and Limitations
|
| 61 |
|
| 62 |
### Quickstart
|
|
|
|
| 160 |
- **LAMBADA**: Text comprehension requiring full-context understanding
|
| 161 |
- **HellaSwag**: Commonsense natural language inference
|
| 162 |
|
| 163 |
+
| Model | Pretraining Filtering | Annealing Filtering | WMDP Bio Average (Robust MCQA, Verified Cloze) (↓) | Average (MMLU, PIQA, Lambada, HellaSwag) (↑) | WMDP Bio Robust MCQA (↓) | WMDP Bio Verified Cloze (↓) | MMLU (↑) | PIQA (↑) | Lambada (↑) | HellaSwag (↑) |
|
| 164 |
+
|:------|:------------------------|:----------------------|:-----------------------------------------------------|:-----------------------------------------------|:---------------------------|:------------------------------|:---------------|:---------------|:---------------|:----------------|
|
| 165 |
+
| deep-ignorance-unfiltered | - | - | 39.66% | 56.05% | 42.97% | 36.34% | 44.92% | 76.44% | 47.08% | 55.75% |
|
| 166 |
+
| deep-ignorance-pretraining-stage-unfiltered | - | - | 37.16% (-2.50) | 60.24% (4.19) | 38.25% (-4.72) | 36.06% (-0.28) | 42.80% (-2.12) | 79.05% (2.61) | 63.03% (15.95) | 56.06% (0.31) |
|
| 167 |
+
| deep-ignorance-e2e-extra-weak-filter | Extra Weak Filter | Extra Weak Filter | 33.70% (-5.96) | 55.83% (-0.22) | 38.02% (-4.95) | 29.37% (-6.97) | 44.13% (-0.79) | 77.04% (0.60) | 46.85% (-0.23) | 55.29% (-0.46) |
|
| 168 |
+
| deep-ignorance-weak-filter-pt-strong-filter-anneal | Weak Filter | Strong Filter | 30.97% (-8.69) | 56.22% (0.17) | 36.75% (-6.22) | 25.19% (-11.15) | 43.16% (-1.76) | 77.20% (0.76) | 48.86% (1.78) | 55.67% (-0.08) |
|
| 169 |
+
| deep-ignorance-e2e-weak-filter | Weak Filter | Weak Filter | 30.50% (-9.16) | 57.37% (1.32) | 35.25% (-7.72) | 25.74% (-10.60) | 43.91% (-1.01) | 78.35% (1.91) | 51.81% (4.73) | 55.41% (-0.34) |
|
| 170 |
+
| deep-ignorance-strong-filter-pt-weak-filter-anneal | Strong Filter | Weak Filter | 30.38% (-9.28) | 57.88% (1.83) | 33.99% (-8.98) | 26.77% (-9.57) | 44.82% (-0.10) | 76.88% (0.44) | 54.05% (6.97) | 55.78% (0.03) |
|
| 171 |
+
| deep-ignorance-e2e-strong-filter | Strong Filter | Strong Filter | 29.90% (-9.76) | 55.53% (-0.52) | 35.37% (-7.60) | 24.44% (-11.90) | 43.21% (-1.71) | 75.73% (-0.71) | 47.29% (0.21) | 55.90% (0.15) |
|
| 172 |
+
| deep-ignorance-pretraining-stage-strong-filter | Strong Filter | - | 29.47% (-10.19) | 60.02% (3.97) | 33.29% (-9.68) | 25.65% (-10.69) | 43.46% (-1.46) | 79.27% (2.83) | 60.82% (13.74) | 56.53% (0.78) |
|
| 173 |
+
| deep-ignorance-unfiltered-cb | - | - | 29.29% (-10.37) | 54.11% (-1.94) | 29.49% (-13.48) | 29.09% (-7.25) | 43.61% (-1.31) | 76.50% (0.06) | 45.84% (-1.24) | 50.50% (-5.25) |
|
| 174 |
+
| deep-ignorance-pretraining-stage-weak-filter | Weak Filter | - | 29.12% (-10.54) | 58.98% (2.93) | 33.53% (-9.44) | 24.72% (-11.62) | 41.04% (-3.88) | 78.78% (2.34) | 60.57% (13.49) | 55.53% (-0.22) |
|
| 175 |
+
| deep-ignorance-strong-filter-pt-weak-filter-anneal-cb-lat | Strong Filter | Weak Filter | 26.92% (-12.74) | 58.00% (1.95) | 29.95% (-13.02) | 23.88% (-12.46) | 43.52% (-1.40) | 76.61% (0.17) | 56.01% (8.93) | 55.84% (0.09) |
|
| 176 |
+
| deep-ignorance-strong-filter-pt-weak-filter-anneal-cb | Strong Filter | Weak Filter | 26.12% (-13.54) | 56.46% (0.41) | 25.46% (-17.51) | 26.77% (-9.57) | 41.45% (-3.47) | 76.33% (-0.11) | 53.64% (6.56) | 54.40% (-1.35) |
|
| 177 |
+
| deep-ignorance-unfiltered-cb-lat | - | - | 25.93% (-13.73) | 56.43% (0.38) | 27.42% (-15.55) | 24.44% (-11.90) | 42.73% (-2.19) | 76.22% (-0.22) | 51.85% (4.77) | 54.92% (-0.83) |
|
| 178 |
+
| deep-ignorance-e2e-strong-filter-cb-lat | Strong Filter | Strong Filter | 25.87% (-13.79) | 56.60% (0.55) | 27.76% (-15.21) | 23.98% (-12.36) | 42.08% (-2.84) | 75.41% (-1.03) | 52.75% (5.67) | 56.18% (0.43) |
|
| 179 |
+
| deep-ignorance-e2e-strong-filter-cb | Strong Filter | Strong Filter | 25.56% (-14.10) | 52.60% (-3.45) | 25.00% (-17.97) | 26.12% (-10.22) | 39.45% (-5.47) | 75.35% (-1.09) | 47.56% (0.48) | 48.03% (-7.72) |
|
| 180 |
|
| 181 |
# Acknowledgments
|
| 182 |
|