jjrussell10 commited on
Commit
b080c5f
·
verified ·
1 Parent(s): 0d96a04

Model cards: cite russell2026wildai; root tokenizer (part 3)

Browse files
This view is limited to 50 files because it contains too many changes.   See raw diff
Files changed (50) hide show
  1. 35.8m-f07-ai-removed/README.md +2 -2
  2. 35.8m-f07-web-mix/README.md +2 -2
  3. 35.8m-f08-ai-removed/README.md +2 -2
  4. 35.8m-f08-web-mix/README.md +2 -2
  5. 35.8m-f09-ai-removed/README.md +2 -2
  6. 35.8m-f09-web-mix/README.md +2 -2
  7. 35.8m-f10-ai-removed/README.md +2 -2
  8. 35.8m-f10-web-mix/README.md +2 -2
  9. 35.8m-g021-ai-r0.05/README.md +2 -2
  10. 35.8m-g021-ai-r0.1/README.md +2 -2
  11. 35.8m-g021-ai-r0.25/README.md +2 -2
  12. 35.8m-g021-ai-r0.5/README.md +2 -2
  13. 35.8m-g021-ai-r1/README.md +2 -2
  14. 35.8m-g021-control/README.md +2 -2
  15. 35.8m-g022-ai-r0.0234/README.md +2 -2
  16. 35.8m-g022-ai-r0.0465/README.md +2 -2
  17. 35.8m-g022-ai-r17.6/README.md +2 -2
  18. 35.8m-g022-ai-r2.1/README.md +2 -2
  19. 35.8m-g022-ai-r35/README.md +2 -2
  20. 35.8m-g022-ai-r4.31/README.md +2 -2
  21. 35.8m-g022-ai-r69.6/README.md +2 -2
  22. 35.8m-g022-ai-r8.76/README.md +2 -2
  23. 35.8m-g022-control/README.md +2 -2
  24. 35.8m-g022-human-r0.025/README.md +2 -2
  25. 35.8m-g022-human-r0.05/README.md +2 -2
  26. 35.8m-g022-human-r4/README.md +2 -2
  27. 35.8m-g023-ai-r0.025/README.md +2 -2
  28. 35.8m-g023-ai-r0.05/README.md +2 -2
  29. 35.8m-g023-ai-r0.1/README.md +2 -2
  30. 35.8m-g023-ai-r0.25/README.md +2 -2
  31. 35.8m-g023-ai-r0.5/README.md +2 -2
  32. 35.8m-g023-ai-r1/README.md +2 -2
  33. 35.8m-g023-ai-r16/README.md +2 -2
  34. 35.8m-g023-ai-r2/README.md +2 -2
  35. 35.8m-g023-ai-r32/README.md +2 -2
  36. 35.8m-g023-ai-r4/README.md +2 -2
  37. 35.8m-g023-ai-r64/README.md +2 -2
  38. 35.8m-g023-ai-r8/README.md +2 -2
  39. 35.8m-g023-control/README.md +2 -2
  40. 35.8m-g023-human-r0.025/README.md +2 -2
  41. 35.8m-g023-human-r0.05/README.md +2 -2
  42. 35.8m-g023-human-r4/README.md +2 -2
  43. 35.8m-g024-ai-r0.05/README.md +2 -2
  44. 35.8m-g024-ai-r0.1/README.md +2 -2
  45. 35.8m-g024-ai-r0.25/README.md +2 -2
  46. 35.8m-g024-ai-r0.5/README.md +2 -2
  47. 35.8m-g024-ai-r1/README.md +2 -2
  48. 35.8m-g024-control/README.md +2 -2
  49. 35.8m-g025-ai-r0.0256/README.md +2 -2
  50. 35.8m-g025-ai-r0.0514/README.md +2 -2
35.8m-f07-ai-removed/README.md CHANGED
@@ -18,7 +18,7 @@ tags:
18
 
19
  > **License: [CC BY-NC-SA 4.0](https://creativecommons.org/licenses/by-nc-sa/4.0/).** The WildAI code, models and dataset are released under the Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International license: you may use and adapt them for non-commercial purposes, with attribution, and must share adaptations under the same license.
20
 
21
- A 35.8M decoder-only language model trained from scratch for [*How Much Is an AI Token Worth? Scaling Laws for Wild AI-Generated Web Text*](https://arxiv.org/abs/2609.40295), one of 844 models that measure how wild AI-generated web text in pretraining data changes a model's loss. It is a research artifact for studying scaling laws, not an assistant: it has had no instruction tuning or safety training.
22
 
23
  ## Training data
24
 
@@ -71,7 +71,7 @@ The nanochat GPT recipe: rotary embeddings, RMSNorm without weights (also on que
71
  ## Citation
72
 
73
  ```bibtex
74
- @article{russell2026aitoken,
75
  title = {How Much Is an AI Token Worth? Scaling Laws for Wild AI-Generated Web Text},
76
  author = {Russell, Jenna and Glickenhaus, Ben and Thai, Katherine and Wieting, John and Iyyer, Mohit and Spero, Max and Emi, Bradley},
77
  journal = {arXiv preprint arXiv:2609.40295},
 
18
 
19
  > **License: [CC BY-NC-SA 4.0](https://creativecommons.org/licenses/by-nc-sa/4.0/).** The WildAI code, models and dataset are released under the Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International license: you may use and adapt them for non-commercial purposes, with attribution, and must share adaptations under the same license.
20
 
21
+ A 35.8M decoder-only language model trained from scratch for [*How Much Is an AI Token Worth? Scaling Laws for Wild AI-Generated Web Text*](https://arxiv.org/abs/2609.40295), one of the paper's models measuring how wild AI-generated web text in pretraining data changes a model's loss. It is a research artifact for studying scaling laws, not an assistant: it has had no instruction tuning or safety training.
22
 
23
  ## Training data
24
 
 
71
  ## Citation
72
 
73
  ```bibtex
74
+ @article{russell2026wildai,
75
  title = {How Much Is an AI Token Worth? Scaling Laws for Wild AI-Generated Web Text},
76
  author = {Russell, Jenna and Glickenhaus, Ben and Thai, Katherine and Wieting, John and Iyyer, Mohit and Spero, Max and Emi, Bradley},
77
  journal = {arXiv preprint arXiv:2609.40295},
35.8m-f07-web-mix/README.md CHANGED
@@ -18,7 +18,7 @@ tags:
18
 
19
  > **License: [CC BY-NC-SA 4.0](https://creativecommons.org/licenses/by-nc-sa/4.0/).** The WildAI code, models and dataset are released under the Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International license: you may use and adapt them for non-commercial purposes, with attribution, and must share adaptations under the same license.
20
 
21
- A 35.8M decoder-only language model trained from scratch for [*How Much Is an AI Token Worth? Scaling Laws for Wild AI-Generated Web Text*](https://arxiv.org/abs/2609.40295), one of 844 models that measure how wild AI-generated web text in pretraining data changes a model's loss. It is a research artifact for studying scaling laws, not an assistant: it has had no instruction tuning or safety training.
22
 
23
  ## Training data
24
 
@@ -71,7 +71,7 @@ The nanochat GPT recipe: rotary embeddings, RMSNorm without weights (also on que
71
  ## Citation
72
 
73
  ```bibtex
74
- @article{russell2026aitoken,
75
  title = {How Much Is an AI Token Worth? Scaling Laws for Wild AI-Generated Web Text},
76
  author = {Russell, Jenna and Glickenhaus, Ben and Thai, Katherine and Wieting, John and Iyyer, Mohit and Spero, Max and Emi, Bradley},
77
  journal = {arXiv preprint arXiv:2609.40295},
 
18
 
19
  > **License: [CC BY-NC-SA 4.0](https://creativecommons.org/licenses/by-nc-sa/4.0/).** The WildAI code, models and dataset are released under the Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International license: you may use and adapt them for non-commercial purposes, with attribution, and must share adaptations under the same license.
20
 
21
+ A 35.8M decoder-only language model trained from scratch for [*How Much Is an AI Token Worth? Scaling Laws for Wild AI-Generated Web Text*](https://arxiv.org/abs/2609.40295), one of the paper's models measuring how wild AI-generated web text in pretraining data changes a model's loss. It is a research artifact for studying scaling laws, not an assistant: it has had no instruction tuning or safety training.
22
 
23
  ## Training data
24
 
 
71
  ## Citation
72
 
73
  ```bibtex
74
+ @article{russell2026wildai,
75
  title = {How Much Is an AI Token Worth? Scaling Laws for Wild AI-Generated Web Text},
76
  author = {Russell, Jenna and Glickenhaus, Ben and Thai, Katherine and Wieting, John and Iyyer, Mohit and Spero, Max and Emi, Bradley},
77
  journal = {arXiv preprint arXiv:2609.40295},
35.8m-f08-ai-removed/README.md CHANGED
@@ -18,7 +18,7 @@ tags:
18
 
19
  > **License: [CC BY-NC-SA 4.0](https://creativecommons.org/licenses/by-nc-sa/4.0/).** The WildAI code, models and dataset are released under the Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International license: you may use and adapt them for non-commercial purposes, with attribution, and must share adaptations under the same license.
20
 
21
- A 35.8M decoder-only language model trained from scratch for [*How Much Is an AI Token Worth? Scaling Laws for Wild AI-Generated Web Text*](https://arxiv.org/abs/2609.40295), one of 844 models that measure how wild AI-generated web text in pretraining data changes a model's loss. It is a research artifact for studying scaling laws, not an assistant: it has had no instruction tuning or safety training.
22
 
23
  ## Training data
24
 
@@ -71,7 +71,7 @@ The nanochat GPT recipe: rotary embeddings, RMSNorm without weights (also on que
71
  ## Citation
72
 
73
  ```bibtex
74
- @article{russell2026aitoken,
75
  title = {How Much Is an AI Token Worth? Scaling Laws for Wild AI-Generated Web Text},
76
  author = {Russell, Jenna and Glickenhaus, Ben and Thai, Katherine and Wieting, John and Iyyer, Mohit and Spero, Max and Emi, Bradley},
77
  journal = {arXiv preprint arXiv:2609.40295},
 
18
 
19
  > **License: [CC BY-NC-SA 4.0](https://creativecommons.org/licenses/by-nc-sa/4.0/).** The WildAI code, models and dataset are released under the Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International license: you may use and adapt them for non-commercial purposes, with attribution, and must share adaptations under the same license.
20
 
21
+ A 35.8M decoder-only language model trained from scratch for [*How Much Is an AI Token Worth? Scaling Laws for Wild AI-Generated Web Text*](https://arxiv.org/abs/2609.40295), one of the paper's models measuring how wild AI-generated web text in pretraining data changes a model's loss. It is a research artifact for studying scaling laws, not an assistant: it has had no instruction tuning or safety training.
22
 
23
  ## Training data
24
 
 
71
  ## Citation
72
 
73
  ```bibtex
74
+ @article{russell2026wildai,
75
  title = {How Much Is an AI Token Worth? Scaling Laws for Wild AI-Generated Web Text},
76
  author = {Russell, Jenna and Glickenhaus, Ben and Thai, Katherine and Wieting, John and Iyyer, Mohit and Spero, Max and Emi, Bradley},
77
  journal = {arXiv preprint arXiv:2609.40295},
35.8m-f08-web-mix/README.md CHANGED
@@ -18,7 +18,7 @@ tags:
18
 
19
  > **License: [CC BY-NC-SA 4.0](https://creativecommons.org/licenses/by-nc-sa/4.0/).** The WildAI code, models and dataset are released under the Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International license: you may use and adapt them for non-commercial purposes, with attribution, and must share adaptations under the same license.
20
 
21
- A 35.8M decoder-only language model trained from scratch for [*How Much Is an AI Token Worth? Scaling Laws for Wild AI-Generated Web Text*](https://arxiv.org/abs/2609.40295), one of 844 models that measure how wild AI-generated web text in pretraining data changes a model's loss. It is a research artifact for studying scaling laws, not an assistant: it has had no instruction tuning or safety training.
22
 
23
  ## Training data
24
 
@@ -71,7 +71,7 @@ The nanochat GPT recipe: rotary embeddings, RMSNorm without weights (also on que
71
  ## Citation
72
 
73
  ```bibtex
74
- @article{russell2026aitoken,
75
  title = {How Much Is an AI Token Worth? Scaling Laws for Wild AI-Generated Web Text},
76
  author = {Russell, Jenna and Glickenhaus, Ben and Thai, Katherine and Wieting, John and Iyyer, Mohit and Spero, Max and Emi, Bradley},
77
  journal = {arXiv preprint arXiv:2609.40295},
 
18
 
19
  > **License: [CC BY-NC-SA 4.0](https://creativecommons.org/licenses/by-nc-sa/4.0/).** The WildAI code, models and dataset are released under the Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International license: you may use and adapt them for non-commercial purposes, with attribution, and must share adaptations under the same license.
20
 
21
+ A 35.8M decoder-only language model trained from scratch for [*How Much Is an AI Token Worth? Scaling Laws for Wild AI-Generated Web Text*](https://arxiv.org/abs/2609.40295), one of the paper's models measuring how wild AI-generated web text in pretraining data changes a model's loss. It is a research artifact for studying scaling laws, not an assistant: it has had no instruction tuning or safety training.
22
 
23
  ## Training data
24
 
 
71
  ## Citation
72
 
73
  ```bibtex
74
+ @article{russell2026wildai,
75
  title = {How Much Is an AI Token Worth? Scaling Laws for Wild AI-Generated Web Text},
76
  author = {Russell, Jenna and Glickenhaus, Ben and Thai, Katherine and Wieting, John and Iyyer, Mohit and Spero, Max and Emi, Bradley},
77
  journal = {arXiv preprint arXiv:2609.40295},
35.8m-f09-ai-removed/README.md CHANGED
@@ -18,7 +18,7 @@ tags:
18
 
19
  > **License: [CC BY-NC-SA 4.0](https://creativecommons.org/licenses/by-nc-sa/4.0/).** The WildAI code, models and dataset are released under the Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International license: you may use and adapt them for non-commercial purposes, with attribution, and must share adaptations under the same license.
20
 
21
- A 35.8M decoder-only language model trained from scratch for [*How Much Is an AI Token Worth? Scaling Laws for Wild AI-Generated Web Text*](https://arxiv.org/abs/2609.40295), one of 844 models that measure how wild AI-generated web text in pretraining data changes a model's loss. It is a research artifact for studying scaling laws, not an assistant: it has had no instruction tuning or safety training.
22
 
23
  ## Training data
24
 
@@ -71,7 +71,7 @@ The nanochat GPT recipe: rotary embeddings, RMSNorm without weights (also on que
71
  ## Citation
72
 
73
  ```bibtex
74
- @article{russell2026aitoken,
75
  title = {How Much Is an AI Token Worth? Scaling Laws for Wild AI-Generated Web Text},
76
  author = {Russell, Jenna and Glickenhaus, Ben and Thai, Katherine and Wieting, John and Iyyer, Mohit and Spero, Max and Emi, Bradley},
77
  journal = {arXiv preprint arXiv:2609.40295},
 
18
 
19
  > **License: [CC BY-NC-SA 4.0](https://creativecommons.org/licenses/by-nc-sa/4.0/).** The WildAI code, models and dataset are released under the Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International license: you may use and adapt them for non-commercial purposes, with attribution, and must share adaptations under the same license.
20
 
21
+ A 35.8M decoder-only language model trained from scratch for [*How Much Is an AI Token Worth? Scaling Laws for Wild AI-Generated Web Text*](https://arxiv.org/abs/2609.40295), one of the paper's models measuring how wild AI-generated web text in pretraining data changes a model's loss. It is a research artifact for studying scaling laws, not an assistant: it has had no instruction tuning or safety training.
22
 
23
  ## Training data
24
 
 
71
  ## Citation
72
 
73
  ```bibtex
74
+ @article{russell2026wildai,
75
  title = {How Much Is an AI Token Worth? Scaling Laws for Wild AI-Generated Web Text},
76
  author = {Russell, Jenna and Glickenhaus, Ben and Thai, Katherine and Wieting, John and Iyyer, Mohit and Spero, Max and Emi, Bradley},
77
  journal = {arXiv preprint arXiv:2609.40295},
35.8m-f09-web-mix/README.md CHANGED
@@ -18,7 +18,7 @@ tags:
18
 
19
  > **License: [CC BY-NC-SA 4.0](https://creativecommons.org/licenses/by-nc-sa/4.0/).** The WildAI code, models and dataset are released under the Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International license: you may use and adapt them for non-commercial purposes, with attribution, and must share adaptations under the same license.
20
 
21
- A 35.8M decoder-only language model trained from scratch for [*How Much Is an AI Token Worth? Scaling Laws for Wild AI-Generated Web Text*](https://arxiv.org/abs/2609.40295), one of 844 models that measure how wild AI-generated web text in pretraining data changes a model's loss. It is a research artifact for studying scaling laws, not an assistant: it has had no instruction tuning or safety training.
22
 
23
  ## Training data
24
 
@@ -71,7 +71,7 @@ The nanochat GPT recipe: rotary embeddings, RMSNorm without weights (also on que
71
  ## Citation
72
 
73
  ```bibtex
74
- @article{russell2026aitoken,
75
  title = {How Much Is an AI Token Worth? Scaling Laws for Wild AI-Generated Web Text},
76
  author = {Russell, Jenna and Glickenhaus, Ben and Thai, Katherine and Wieting, John and Iyyer, Mohit and Spero, Max and Emi, Bradley},
77
  journal = {arXiv preprint arXiv:2609.40295},
 
18
 
19
  > **License: [CC BY-NC-SA 4.0](https://creativecommons.org/licenses/by-nc-sa/4.0/).** The WildAI code, models and dataset are released under the Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International license: you may use and adapt them for non-commercial purposes, with attribution, and must share adaptations under the same license.
20
 
21
+ A 35.8M decoder-only language model trained from scratch for [*How Much Is an AI Token Worth? Scaling Laws for Wild AI-Generated Web Text*](https://arxiv.org/abs/2609.40295), one of the paper's models measuring how wild AI-generated web text in pretraining data changes a model's loss. It is a research artifact for studying scaling laws, not an assistant: it has had no instruction tuning or safety training.
22
 
23
  ## Training data
24
 
 
71
  ## Citation
72
 
73
  ```bibtex
74
+ @article{russell2026wildai,
75
  title = {How Much Is an AI Token Worth? Scaling Laws for Wild AI-Generated Web Text},
76
  author = {Russell, Jenna and Glickenhaus, Ben and Thai, Katherine and Wieting, John and Iyyer, Mohit and Spero, Max and Emi, Bradley},
77
  journal = {arXiv preprint arXiv:2609.40295},
35.8m-f10-ai-removed/README.md CHANGED
@@ -18,7 +18,7 @@ tags:
18
 
19
  > **License: [CC BY-NC-SA 4.0](https://creativecommons.org/licenses/by-nc-sa/4.0/).** The WildAI code, models and dataset are released under the Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International license: you may use and adapt them for non-commercial purposes, with attribution, and must share adaptations under the same license.
20
 
21
- A 35.8M decoder-only language model trained from scratch for [*How Much Is an AI Token Worth? Scaling Laws for Wild AI-Generated Web Text*](https://arxiv.org/abs/2609.40295), one of 844 models that measure how wild AI-generated web text in pretraining data changes a model's loss. It is a research artifact for studying scaling laws, not an assistant: it has had no instruction tuning or safety training.
22
 
23
  ## Training data
24
 
@@ -71,7 +71,7 @@ The nanochat GPT recipe: rotary embeddings, RMSNorm without weights (also on que
71
  ## Citation
72
 
73
  ```bibtex
74
- @article{russell2026aitoken,
75
  title = {How Much Is an AI Token Worth? Scaling Laws for Wild AI-Generated Web Text},
76
  author = {Russell, Jenna and Glickenhaus, Ben and Thai, Katherine and Wieting, John and Iyyer, Mohit and Spero, Max and Emi, Bradley},
77
  journal = {arXiv preprint arXiv:2609.40295},
 
18
 
19
  > **License: [CC BY-NC-SA 4.0](https://creativecommons.org/licenses/by-nc-sa/4.0/).** The WildAI code, models and dataset are released under the Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International license: you may use and adapt them for non-commercial purposes, with attribution, and must share adaptations under the same license.
20
 
21
+ A 35.8M decoder-only language model trained from scratch for [*How Much Is an AI Token Worth? Scaling Laws for Wild AI-Generated Web Text*](https://arxiv.org/abs/2609.40295), one of the paper's models measuring how wild AI-generated web text in pretraining data changes a model's loss. It is a research artifact for studying scaling laws, not an assistant: it has had no instruction tuning or safety training.
22
 
23
  ## Training data
24
 
 
71
  ## Citation
72
 
73
  ```bibtex
74
+ @article{russell2026wildai,
75
  title = {How Much Is an AI Token Worth? Scaling Laws for Wild AI-Generated Web Text},
76
  author = {Russell, Jenna and Glickenhaus, Ben and Thai, Katherine and Wieting, John and Iyyer, Mohit and Spero, Max and Emi, Bradley},
77
  journal = {arXiv preprint arXiv:2609.40295},
35.8m-f10-web-mix/README.md CHANGED
@@ -18,7 +18,7 @@ tags:
18
 
19
  > **License: [CC BY-NC-SA 4.0](https://creativecommons.org/licenses/by-nc-sa/4.0/).** The WildAI code, models and dataset are released under the Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International license: you may use and adapt them for non-commercial purposes, with attribution, and must share adaptations under the same license.
20
 
21
- A 35.8M decoder-only language model trained from scratch for [*How Much Is an AI Token Worth? Scaling Laws for Wild AI-Generated Web Text*](https://arxiv.org/abs/2609.40295), one of 844 models that measure how wild AI-generated web text in pretraining data changes a model's loss. It is a research artifact for studying scaling laws, not an assistant: it has had no instruction tuning or safety training.
22
 
23
  ## Training data
24
 
@@ -71,7 +71,7 @@ The nanochat GPT recipe: rotary embeddings, RMSNorm without weights (also on que
71
  ## Citation
72
 
73
  ```bibtex
74
- @article{russell2026aitoken,
75
  title = {How Much Is an AI Token Worth? Scaling Laws for Wild AI-Generated Web Text},
76
  author = {Russell, Jenna and Glickenhaus, Ben and Thai, Katherine and Wieting, John and Iyyer, Mohit and Spero, Max and Emi, Bradley},
77
  journal = {arXiv preprint arXiv:2609.40295},
 
18
 
19
  > **License: [CC BY-NC-SA 4.0](https://creativecommons.org/licenses/by-nc-sa/4.0/).** The WildAI code, models and dataset are released under the Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International license: you may use and adapt them for non-commercial purposes, with attribution, and must share adaptations under the same license.
20
 
21
+ A 35.8M decoder-only language model trained from scratch for [*How Much Is an AI Token Worth? Scaling Laws for Wild AI-Generated Web Text*](https://arxiv.org/abs/2609.40295), one of the paper's models measuring how wild AI-generated web text in pretraining data changes a model's loss. It is a research artifact for studying scaling laws, not an assistant: it has had no instruction tuning or safety training.
22
 
23
  ## Training data
24
 
 
71
  ## Citation
72
 
73
  ```bibtex
74
+ @article{russell2026wildai,
75
  title = {How Much Is an AI Token Worth? Scaling Laws for Wild AI-Generated Web Text},
76
  author = {Russell, Jenna and Glickenhaus, Ben and Thai, Katherine and Wieting, John and Iyyer, Mohit and Spero, Max and Emi, Bradley},
77
  journal = {arXiv preprint arXiv:2609.40295},
35.8m-g021-ai-r0.05/README.md CHANGED
@@ -18,7 +18,7 @@ tags:
18
 
19
  > **License: [CC BY-NC-SA 4.0](https://creativecommons.org/licenses/by-nc-sa/4.0/).** The WildAI code, models and dataset are released under the Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International license: you may use and adapt them for non-commercial purposes, with attribution, and must share adaptations under the same license.
20
 
21
- A 35.8M decoder-only language model trained from scratch for [*How Much Is an AI Token Worth? Scaling Laws for Wild AI-Generated Web Text*](https://arxiv.org/abs/2609.40295), one of 844 models that measure how wild AI-generated web text in pretraining data changes a model's loss. It is a research artifact for studying scaling laws, not an assistant: it has had no instruction tuning or safety training.
22
 
23
  ## Training data
24
 
@@ -95,7 +95,7 @@ The nanochat GPT recipe: rotary embeddings, RMSNorm without weights (also on que
95
  ## Citation
96
 
97
  ```bibtex
98
- @article{russell2026aitoken,
99
  title = {How Much Is an AI Token Worth? Scaling Laws for Wild AI-Generated Web Text},
100
  author = {Russell, Jenna and Glickenhaus, Ben and Thai, Katherine and Wieting, John and Iyyer, Mohit and Spero, Max and Emi, Bradley},
101
  journal = {arXiv preprint arXiv:2609.40295},
 
18
 
19
  > **License: [CC BY-NC-SA 4.0](https://creativecommons.org/licenses/by-nc-sa/4.0/).** The WildAI code, models and dataset are released under the Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International license: you may use and adapt them for non-commercial purposes, with attribution, and must share adaptations under the same license.
20
 
21
+ A 35.8M decoder-only language model trained from scratch for [*How Much Is an AI Token Worth? Scaling Laws for Wild AI-Generated Web Text*](https://arxiv.org/abs/2609.40295), one of the paper's models measuring how wild AI-generated web text in pretraining data changes a model's loss. It is a research artifact for studying scaling laws, not an assistant: it has had no instruction tuning or safety training.
22
 
23
  ## Training data
24
 
 
95
  ## Citation
96
 
97
  ```bibtex
98
+ @article{russell2026wildai,
99
  title = {How Much Is an AI Token Worth? Scaling Laws for Wild AI-Generated Web Text},
100
  author = {Russell, Jenna and Glickenhaus, Ben and Thai, Katherine and Wieting, John and Iyyer, Mohit and Spero, Max and Emi, Bradley},
101
  journal = {arXiv preprint arXiv:2609.40295},
35.8m-g021-ai-r0.1/README.md CHANGED
@@ -18,7 +18,7 @@ tags:
18
 
19
  > **License: [CC BY-NC-SA 4.0](https://creativecommons.org/licenses/by-nc-sa/4.0/).** The WildAI code, models and dataset are released under the Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International license: you may use and adapt them for non-commercial purposes, with attribution, and must share adaptations under the same license.
20
 
21
- A 35.8M decoder-only language model trained from scratch for [*How Much Is an AI Token Worth? Scaling Laws for Wild AI-Generated Web Text*](https://arxiv.org/abs/2609.40295), one of 844 models that measure how wild AI-generated web text in pretraining data changes a model's loss. It is a research artifact for studying scaling laws, not an assistant: it has had no instruction tuning or safety training.
22
 
23
  ## Training data
24
 
@@ -95,7 +95,7 @@ The nanochat GPT recipe: rotary embeddings, RMSNorm without weights (also on que
95
  ## Citation
96
 
97
  ```bibtex
98
- @article{russell2026aitoken,
99
  title = {How Much Is an AI Token Worth? Scaling Laws for Wild AI-Generated Web Text},
100
  author = {Russell, Jenna and Glickenhaus, Ben and Thai, Katherine and Wieting, John and Iyyer, Mohit and Spero, Max and Emi, Bradley},
101
  journal = {arXiv preprint arXiv:2609.40295},
 
18
 
19
  > **License: [CC BY-NC-SA 4.0](https://creativecommons.org/licenses/by-nc-sa/4.0/).** The WildAI code, models and dataset are released under the Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International license: you may use and adapt them for non-commercial purposes, with attribution, and must share adaptations under the same license.
20
 
21
+ A 35.8M decoder-only language model trained from scratch for [*How Much Is an AI Token Worth? Scaling Laws for Wild AI-Generated Web Text*](https://arxiv.org/abs/2609.40295), one of the paper's models measuring how wild AI-generated web text in pretraining data changes a model's loss. It is a research artifact for studying scaling laws, not an assistant: it has had no instruction tuning or safety training.
22
 
23
  ## Training data
24
 
 
95
  ## Citation
96
 
97
  ```bibtex
98
+ @article{russell2026wildai,
99
  title = {How Much Is an AI Token Worth? Scaling Laws for Wild AI-Generated Web Text},
100
  author = {Russell, Jenna and Glickenhaus, Ben and Thai, Katherine and Wieting, John and Iyyer, Mohit and Spero, Max and Emi, Bradley},
101
  journal = {arXiv preprint arXiv:2609.40295},
35.8m-g021-ai-r0.25/README.md CHANGED
@@ -18,7 +18,7 @@ tags:
18
 
19
  > **License: [CC BY-NC-SA 4.0](https://creativecommons.org/licenses/by-nc-sa/4.0/).** The WildAI code, models and dataset are released under the Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International license: you may use and adapt them for non-commercial purposes, with attribution, and must share adaptations under the same license.
20
 
21
- A 35.8M decoder-only language model trained from scratch for [*How Much Is an AI Token Worth? Scaling Laws for Wild AI-Generated Web Text*](https://arxiv.org/abs/2609.40295), one of 844 models that measure how wild AI-generated web text in pretraining data changes a model's loss. It is a research artifact for studying scaling laws, not an assistant: it has had no instruction tuning or safety training.
22
 
23
  ## Training data
24
 
@@ -95,7 +95,7 @@ The nanochat GPT recipe: rotary embeddings, RMSNorm without weights (also on que
95
  ## Citation
96
 
97
  ```bibtex
98
- @article{russell2026aitoken,
99
  title = {How Much Is an AI Token Worth? Scaling Laws for Wild AI-Generated Web Text},
100
  author = {Russell, Jenna and Glickenhaus, Ben and Thai, Katherine and Wieting, John and Iyyer, Mohit and Spero, Max and Emi, Bradley},
101
  journal = {arXiv preprint arXiv:2609.40295},
 
18
 
19
  > **License: [CC BY-NC-SA 4.0](https://creativecommons.org/licenses/by-nc-sa/4.0/).** The WildAI code, models and dataset are released under the Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International license: you may use and adapt them for non-commercial purposes, with attribution, and must share adaptations under the same license.
20
 
21
+ A 35.8M decoder-only language model trained from scratch for [*How Much Is an AI Token Worth? Scaling Laws for Wild AI-Generated Web Text*](https://arxiv.org/abs/2609.40295), one of the paper's models measuring how wild AI-generated web text in pretraining data changes a model's loss. It is a research artifact for studying scaling laws, not an assistant: it has had no instruction tuning or safety training.
22
 
23
  ## Training data
24
 
 
95
  ## Citation
96
 
97
  ```bibtex
98
+ @article{russell2026wildai,
99
  title = {How Much Is an AI Token Worth? Scaling Laws for Wild AI-Generated Web Text},
100
  author = {Russell, Jenna and Glickenhaus, Ben and Thai, Katherine and Wieting, John and Iyyer, Mohit and Spero, Max and Emi, Bradley},
101
  journal = {arXiv preprint arXiv:2609.40295},
35.8m-g021-ai-r0.5/README.md CHANGED
@@ -18,7 +18,7 @@ tags:
18
 
19
  > **License: [CC BY-NC-SA 4.0](https://creativecommons.org/licenses/by-nc-sa/4.0/).** The WildAI code, models and dataset are released under the Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International license: you may use and adapt them for non-commercial purposes, with attribution, and must share adaptations under the same license.
20
 
21
- A 35.8M decoder-only language model trained from scratch for [*How Much Is an AI Token Worth? Scaling Laws for Wild AI-Generated Web Text*](https://arxiv.org/abs/2609.40295), one of 844 models that measure how wild AI-generated web text in pretraining data changes a model's loss. It is a research artifact for studying scaling laws, not an assistant: it has had no instruction tuning or safety training.
22
 
23
  ## Training data
24
 
@@ -95,7 +95,7 @@ The nanochat GPT recipe: rotary embeddings, RMSNorm without weights (also on que
95
  ## Citation
96
 
97
  ```bibtex
98
- @article{russell2026aitoken,
99
  title = {How Much Is an AI Token Worth? Scaling Laws for Wild AI-Generated Web Text},
100
  author = {Russell, Jenna and Glickenhaus, Ben and Thai, Katherine and Wieting, John and Iyyer, Mohit and Spero, Max and Emi, Bradley},
101
  journal = {arXiv preprint arXiv:2609.40295},
 
18
 
19
  > **License: [CC BY-NC-SA 4.0](https://creativecommons.org/licenses/by-nc-sa/4.0/).** The WildAI code, models and dataset are released under the Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International license: you may use and adapt them for non-commercial purposes, with attribution, and must share adaptations under the same license.
20
 
21
+ A 35.8M decoder-only language model trained from scratch for [*How Much Is an AI Token Worth? Scaling Laws for Wild AI-Generated Web Text*](https://arxiv.org/abs/2609.40295), one of the paper's models measuring how wild AI-generated web text in pretraining data changes a model's loss. It is a research artifact for studying scaling laws, not an assistant: it has had no instruction tuning or safety training.
22
 
23
  ## Training data
24
 
 
95
  ## Citation
96
 
97
  ```bibtex
98
+ @article{russell2026wildai,
99
  title = {How Much Is an AI Token Worth? Scaling Laws for Wild AI-Generated Web Text},
100
  author = {Russell, Jenna and Glickenhaus, Ben and Thai, Katherine and Wieting, John and Iyyer, Mohit and Spero, Max and Emi, Bradley},
101
  journal = {arXiv preprint arXiv:2609.40295},
35.8m-g021-ai-r1/README.md CHANGED
@@ -18,7 +18,7 @@ tags:
18
 
19
  > **License: [CC BY-NC-SA 4.0](https://creativecommons.org/licenses/by-nc-sa/4.0/).** The WildAI code, models and dataset are released under the Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International license: you may use and adapt them for non-commercial purposes, with attribution, and must share adaptations under the same license.
20
 
21
- A 35.8M decoder-only language model trained from scratch for [*How Much Is an AI Token Worth? Scaling Laws for Wild AI-Generated Web Text*](https://arxiv.org/abs/2609.40295), one of 844 models that measure how wild AI-generated web text in pretraining data changes a model's loss. It is a research artifact for studying scaling laws, not an assistant: it has had no instruction tuning or safety training.
22
 
23
  ## Training data
24
 
@@ -95,7 +95,7 @@ The nanochat GPT recipe: rotary embeddings, RMSNorm without weights (also on que
95
  ## Citation
96
 
97
  ```bibtex
98
- @article{russell2026aitoken,
99
  title = {How Much Is an AI Token Worth? Scaling Laws for Wild AI-Generated Web Text},
100
  author = {Russell, Jenna and Glickenhaus, Ben and Thai, Katherine and Wieting, John and Iyyer, Mohit and Spero, Max and Emi, Bradley},
101
  journal = {arXiv preprint arXiv:2609.40295},
 
18
 
19
  > **License: [CC BY-NC-SA 4.0](https://creativecommons.org/licenses/by-nc-sa/4.0/).** The WildAI code, models and dataset are released under the Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International license: you may use and adapt them for non-commercial purposes, with attribution, and must share adaptations under the same license.
20
 
21
+ A 35.8M decoder-only language model trained from scratch for [*How Much Is an AI Token Worth? Scaling Laws for Wild AI-Generated Web Text*](https://arxiv.org/abs/2609.40295), one of the paper's models measuring how wild AI-generated web text in pretraining data changes a model's loss. It is a research artifact for studying scaling laws, not an assistant: it has had no instruction tuning or safety training.
22
 
23
  ## Training data
24
 
 
95
  ## Citation
96
 
97
  ```bibtex
98
+ @article{russell2026wildai,
99
  title = {How Much Is an AI Token Worth? Scaling Laws for Wild AI-Generated Web Text},
100
  author = {Russell, Jenna and Glickenhaus, Ben and Thai, Katherine and Wieting, John and Iyyer, Mohit and Spero, Max and Emi, Bradley},
101
  journal = {arXiv preprint arXiv:2609.40295},
35.8m-g021-control/README.md CHANGED
@@ -18,7 +18,7 @@ tags:
18
 
19
  > **License: [CC BY-NC-SA 4.0](https://creativecommons.org/licenses/by-nc-sa/4.0/).** The WildAI code, models and dataset are released under the Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International license: you may use and adapt them for non-commercial purposes, with attribution, and must share adaptations under the same license.
20
 
21
- A 35.8M decoder-only language model trained from scratch for [*How Much Is an AI Token Worth? Scaling Laws for Wild AI-Generated Web Text*](https://arxiv.org/abs/2609.40295), one of 844 models that measure how wild AI-generated web text in pretraining data changes a model's loss. It is a research artifact for studying scaling laws, not an assistant: it has had no instruction tuning or safety training.
22
 
23
  ## Training data
24
 
@@ -95,7 +95,7 @@ The nanochat GPT recipe: rotary embeddings, RMSNorm without weights (also on que
95
  ## Citation
96
 
97
  ```bibtex
98
- @article{russell2026aitoken,
99
  title = {How Much Is an AI Token Worth? Scaling Laws for Wild AI-Generated Web Text},
100
  author = {Russell, Jenna and Glickenhaus, Ben and Thai, Katherine and Wieting, John and Iyyer, Mohit and Spero, Max and Emi, Bradley},
101
  journal = {arXiv preprint arXiv:2609.40295},
 
18
 
19
  > **License: [CC BY-NC-SA 4.0](https://creativecommons.org/licenses/by-nc-sa/4.0/).** The WildAI code, models and dataset are released under the Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International license: you may use and adapt them for non-commercial purposes, with attribution, and must share adaptations under the same license.
20
 
21
+ A 35.8M decoder-only language model trained from scratch for [*How Much Is an AI Token Worth? Scaling Laws for Wild AI-Generated Web Text*](https://arxiv.org/abs/2609.40295), one of the paper's models measuring how wild AI-generated web text in pretraining data changes a model's loss. It is a research artifact for studying scaling laws, not an assistant: it has had no instruction tuning or safety training.
22
 
23
  ## Training data
24
 
 
95
  ## Citation
96
 
97
  ```bibtex
98
+ @article{russell2026wildai,
99
  title = {How Much Is an AI Token Worth? Scaling Laws for Wild AI-Generated Web Text},
100
  author = {Russell, Jenna and Glickenhaus, Ben and Thai, Katherine and Wieting, John and Iyyer, Mohit and Spero, Max and Emi, Bradley},
101
  journal = {arXiv preprint arXiv:2609.40295},
35.8m-g022-ai-r0.0234/README.md CHANGED
@@ -18,7 +18,7 @@ tags:
18
 
19
  > **License: [CC BY-NC-SA 4.0](https://creativecommons.org/licenses/by-nc-sa/4.0/).** The WildAI code, models and dataset are released under the Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International license: you may use and adapt them for non-commercial purposes, with attribution, and must share adaptations under the same license.
20
 
21
- A 35.8M decoder-only language model trained from scratch for [*How Much Is an AI Token Worth? Scaling Laws for Wild AI-Generated Web Text*](https://arxiv.org/abs/2609.40295), one of 844 models that measure how wild AI-generated web text in pretraining data changes a model's loss. It is a research artifact for studying scaling laws, not an assistant: it has had no instruction tuning or safety training.
22
 
23
  ## Training data
24
 
@@ -95,7 +95,7 @@ The nanochat GPT recipe: rotary embeddings, RMSNorm without weights (also on que
95
  ## Citation
96
 
97
  ```bibtex
98
- @article{russell2026aitoken,
99
  title = {How Much Is an AI Token Worth? Scaling Laws for Wild AI-Generated Web Text},
100
  author = {Russell, Jenna and Glickenhaus, Ben and Thai, Katherine and Wieting, John and Iyyer, Mohit and Spero, Max and Emi, Bradley},
101
  journal = {arXiv preprint arXiv:2609.40295},
 
18
 
19
  > **License: [CC BY-NC-SA 4.0](https://creativecommons.org/licenses/by-nc-sa/4.0/).** The WildAI code, models and dataset are released under the Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International license: you may use and adapt them for non-commercial purposes, with attribution, and must share adaptations under the same license.
20
 
21
+ A 35.8M decoder-only language model trained from scratch for [*How Much Is an AI Token Worth? Scaling Laws for Wild AI-Generated Web Text*](https://arxiv.org/abs/2609.40295), one of the paper's models measuring how wild AI-generated web text in pretraining data changes a model's loss. It is a research artifact for studying scaling laws, not an assistant: it has had no instruction tuning or safety training.
22
 
23
  ## Training data
24
 
 
95
  ## Citation
96
 
97
  ```bibtex
98
+ @article{russell2026wildai,
99
  title = {How Much Is an AI Token Worth? Scaling Laws for Wild AI-Generated Web Text},
100
  author = {Russell, Jenna and Glickenhaus, Ben and Thai, Katherine and Wieting, John and Iyyer, Mohit and Spero, Max and Emi, Bradley},
101
  journal = {arXiv preprint arXiv:2609.40295},
35.8m-g022-ai-r0.0465/README.md CHANGED
@@ -18,7 +18,7 @@ tags:
18
 
19
  > **License: [CC BY-NC-SA 4.0](https://creativecommons.org/licenses/by-nc-sa/4.0/).** The WildAI code, models and dataset are released under the Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International license: you may use and adapt them for non-commercial purposes, with attribution, and must share adaptations under the same license.
20
 
21
- A 35.8M decoder-only language model trained from scratch for [*How Much Is an AI Token Worth? Scaling Laws for Wild AI-Generated Web Text*](https://arxiv.org/abs/2609.40295), one of 844 models that measure how wild AI-generated web text in pretraining data changes a model's loss. It is a research artifact for studying scaling laws, not an assistant: it has had no instruction tuning or safety training.
22
 
23
  ## Training data
24
 
@@ -95,7 +95,7 @@ The nanochat GPT recipe: rotary embeddings, RMSNorm without weights (also on que
95
  ## Citation
96
 
97
  ```bibtex
98
- @article{russell2026aitoken,
99
  title = {How Much Is an AI Token Worth? Scaling Laws for Wild AI-Generated Web Text},
100
  author = {Russell, Jenna and Glickenhaus, Ben and Thai, Katherine and Wieting, John and Iyyer, Mohit and Spero, Max and Emi, Bradley},
101
  journal = {arXiv preprint arXiv:2609.40295},
 
18
 
19
  > **License: [CC BY-NC-SA 4.0](https://creativecommons.org/licenses/by-nc-sa/4.0/).** The WildAI code, models and dataset are released under the Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International license: you may use and adapt them for non-commercial purposes, with attribution, and must share adaptations under the same license.
20
 
21
+ A 35.8M decoder-only language model trained from scratch for [*How Much Is an AI Token Worth? Scaling Laws for Wild AI-Generated Web Text*](https://arxiv.org/abs/2609.40295), one of the paper's models measuring how wild AI-generated web text in pretraining data changes a model's loss. It is a research artifact for studying scaling laws, not an assistant: it has had no instruction tuning or safety training.
22
 
23
  ## Training data
24
 
 
95
  ## Citation
96
 
97
  ```bibtex
98
+ @article{russell2026wildai,
99
  title = {How Much Is an AI Token Worth? Scaling Laws for Wild AI-Generated Web Text},
100
  author = {Russell, Jenna and Glickenhaus, Ben and Thai, Katherine and Wieting, John and Iyyer, Mohit and Spero, Max and Emi, Bradley},
101
  journal = {arXiv preprint arXiv:2609.40295},
35.8m-g022-ai-r17.6/README.md CHANGED
@@ -18,7 +18,7 @@ tags:
18
 
19
  > **License: [CC BY-NC-SA 4.0](https://creativecommons.org/licenses/by-nc-sa/4.0/).** The WildAI code, models and dataset are released under the Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International license: you may use and adapt them for non-commercial purposes, with attribution, and must share adaptations under the same license.
20
 
21
- A 35.8M decoder-only language model trained from scratch for [*How Much Is an AI Token Worth? Scaling Laws for Wild AI-Generated Web Text*](https://arxiv.org/abs/2609.40295), one of 844 models that measure how wild AI-generated web text in pretraining data changes a model's loss. It is a research artifact for studying scaling laws, not an assistant: it has had no instruction tuning or safety training.
22
 
23
  ## Training data
24
 
@@ -95,7 +95,7 @@ The nanochat GPT recipe: rotary embeddings, RMSNorm without weights (also on que
95
  ## Citation
96
 
97
  ```bibtex
98
- @article{russell2026aitoken,
99
  title = {How Much Is an AI Token Worth? Scaling Laws for Wild AI-Generated Web Text},
100
  author = {Russell, Jenna and Glickenhaus, Ben and Thai, Katherine and Wieting, John and Iyyer, Mohit and Spero, Max and Emi, Bradley},
101
  journal = {arXiv preprint arXiv:2609.40295},
 
18
 
19
  > **License: [CC BY-NC-SA 4.0](https://creativecommons.org/licenses/by-nc-sa/4.0/).** The WildAI code, models and dataset are released under the Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International license: you may use and adapt them for non-commercial purposes, with attribution, and must share adaptations under the same license.
20
 
21
+ A 35.8M decoder-only language model trained from scratch for [*How Much Is an AI Token Worth? Scaling Laws for Wild AI-Generated Web Text*](https://arxiv.org/abs/2609.40295), one of the paper's models measuring how wild AI-generated web text in pretraining data changes a model's loss. It is a research artifact for studying scaling laws, not an assistant: it has had no instruction tuning or safety training.
22
 
23
  ## Training data
24
 
 
95
  ## Citation
96
 
97
  ```bibtex
98
+ @article{russell2026wildai,
99
  title = {How Much Is an AI Token Worth? Scaling Laws for Wild AI-Generated Web Text},
100
  author = {Russell, Jenna and Glickenhaus, Ben and Thai, Katherine and Wieting, John and Iyyer, Mohit and Spero, Max and Emi, Bradley},
101
  journal = {arXiv preprint arXiv:2609.40295},
35.8m-g022-ai-r2.1/README.md CHANGED
@@ -18,7 +18,7 @@ tags:
18
 
19
  > **License: [CC BY-NC-SA 4.0](https://creativecommons.org/licenses/by-nc-sa/4.0/).** The WildAI code, models and dataset are released under the Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International license: you may use and adapt them for non-commercial purposes, with attribution, and must share adaptations under the same license.
20
 
21
- A 35.8M decoder-only language model trained from scratch for [*How Much Is an AI Token Worth? Scaling Laws for Wild AI-Generated Web Text*](https://arxiv.org/abs/2609.40295), one of 844 models that measure how wild AI-generated web text in pretraining data changes a model's loss. It is a research artifact for studying scaling laws, not an assistant: it has had no instruction tuning or safety training.
22
 
23
  ## Training data
24
 
@@ -95,7 +95,7 @@ The nanochat GPT recipe: rotary embeddings, RMSNorm without weights (also on que
95
  ## Citation
96
 
97
  ```bibtex
98
- @article{russell2026aitoken,
99
  title = {How Much Is an AI Token Worth? Scaling Laws for Wild AI-Generated Web Text},
100
  author = {Russell, Jenna and Glickenhaus, Ben and Thai, Katherine and Wieting, John and Iyyer, Mohit and Spero, Max and Emi, Bradley},
101
  journal = {arXiv preprint arXiv:2609.40295},
 
18
 
19
  > **License: [CC BY-NC-SA 4.0](https://creativecommons.org/licenses/by-nc-sa/4.0/).** The WildAI code, models and dataset are released under the Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International license: you may use and adapt them for non-commercial purposes, with attribution, and must share adaptations under the same license.
20
 
21
+ A 35.8M decoder-only language model trained from scratch for [*How Much Is an AI Token Worth? Scaling Laws for Wild AI-Generated Web Text*](https://arxiv.org/abs/2609.40295), one of the paper's models measuring how wild AI-generated web text in pretraining data changes a model's loss. It is a research artifact for studying scaling laws, not an assistant: it has had no instruction tuning or safety training.
22
 
23
  ## Training data
24
 
 
95
  ## Citation
96
 
97
  ```bibtex
98
+ @article{russell2026wildai,
99
  title = {How Much Is an AI Token Worth? Scaling Laws for Wild AI-Generated Web Text},
100
  author = {Russell, Jenna and Glickenhaus, Ben and Thai, Katherine and Wieting, John and Iyyer, Mohit and Spero, Max and Emi, Bradley},
101
  journal = {arXiv preprint arXiv:2609.40295},
35.8m-g022-ai-r35/README.md CHANGED
@@ -18,7 +18,7 @@ tags:
18
 
19
  > **License: [CC BY-NC-SA 4.0](https://creativecommons.org/licenses/by-nc-sa/4.0/).** The WildAI code, models and dataset are released under the Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International license: you may use and adapt them for non-commercial purposes, with attribution, and must share adaptations under the same license.
20
 
21
- A 35.8M decoder-only language model trained from scratch for [*How Much Is an AI Token Worth? Scaling Laws for Wild AI-Generated Web Text*](https://arxiv.org/abs/2609.40295), one of 844 models that measure how wild AI-generated web text in pretraining data changes a model's loss. It is a research artifact for studying scaling laws, not an assistant: it has had no instruction tuning or safety training.
22
 
23
  ## Training data
24
 
@@ -95,7 +95,7 @@ The nanochat GPT recipe: rotary embeddings, RMSNorm without weights (also on que
95
  ## Citation
96
 
97
  ```bibtex
98
- @article{russell2026aitoken,
99
  title = {How Much Is an AI Token Worth? Scaling Laws for Wild AI-Generated Web Text},
100
  author = {Russell, Jenna and Glickenhaus, Ben and Thai, Katherine and Wieting, John and Iyyer, Mohit and Spero, Max and Emi, Bradley},
101
  journal = {arXiv preprint arXiv:2609.40295},
 
18
 
19
  > **License: [CC BY-NC-SA 4.0](https://creativecommons.org/licenses/by-nc-sa/4.0/).** The WildAI code, models and dataset are released under the Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International license: you may use and adapt them for non-commercial purposes, with attribution, and must share adaptations under the same license.
20
 
21
+ A 35.8M decoder-only language model trained from scratch for [*How Much Is an AI Token Worth? Scaling Laws for Wild AI-Generated Web Text*](https://arxiv.org/abs/2609.40295), one of the paper's models measuring how wild AI-generated web text in pretraining data changes a model's loss. It is a research artifact for studying scaling laws, not an assistant: it has had no instruction tuning or safety training.
22
 
23
  ## Training data
24
 
 
95
  ## Citation
96
 
97
  ```bibtex
98
+ @article{russell2026wildai,
99
  title = {How Much Is an AI Token Worth? Scaling Laws for Wild AI-Generated Web Text},
100
  author = {Russell, Jenna and Glickenhaus, Ben and Thai, Katherine and Wieting, John and Iyyer, Mohit and Spero, Max and Emi, Bradley},
101
  journal = {arXiv preprint arXiv:2609.40295},
35.8m-g022-ai-r4.31/README.md CHANGED
@@ -18,7 +18,7 @@ tags:
18
 
19
  > **License: [CC BY-NC-SA 4.0](https://creativecommons.org/licenses/by-nc-sa/4.0/).** The WildAI code, models and dataset are released under the Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International license: you may use and adapt them for non-commercial purposes, with attribution, and must share adaptations under the same license.
20
 
21
- A 35.8M decoder-only language model trained from scratch for [*How Much Is an AI Token Worth? Scaling Laws for Wild AI-Generated Web Text*](https://arxiv.org/abs/2609.40295), one of 844 models that measure how wild AI-generated web text in pretraining data changes a model's loss. It is a research artifact for studying scaling laws, not an assistant: it has had no instruction tuning or safety training.
22
 
23
  ## Training data
24
 
@@ -95,7 +95,7 @@ The nanochat GPT recipe: rotary embeddings, RMSNorm without weights (also on que
95
  ## Citation
96
 
97
  ```bibtex
98
- @article{russell2026aitoken,
99
  title = {How Much Is an AI Token Worth? Scaling Laws for Wild AI-Generated Web Text},
100
  author = {Russell, Jenna and Glickenhaus, Ben and Thai, Katherine and Wieting, John and Iyyer, Mohit and Spero, Max and Emi, Bradley},
101
  journal = {arXiv preprint arXiv:2609.40295},
 
18
 
19
  > **License: [CC BY-NC-SA 4.0](https://creativecommons.org/licenses/by-nc-sa/4.0/).** The WildAI code, models and dataset are released under the Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International license: you may use and adapt them for non-commercial purposes, with attribution, and must share adaptations under the same license.
20
 
21
+ A 35.8M decoder-only language model trained from scratch for [*How Much Is an AI Token Worth? Scaling Laws for Wild AI-Generated Web Text*](https://arxiv.org/abs/2609.40295), one of the paper's models measuring how wild AI-generated web text in pretraining data changes a model's loss. It is a research artifact for studying scaling laws, not an assistant: it has had no instruction tuning or safety training.
22
 
23
  ## Training data
24
 
 
95
  ## Citation
96
 
97
  ```bibtex
98
+ @article{russell2026wildai,
99
  title = {How Much Is an AI Token Worth? Scaling Laws for Wild AI-Generated Web Text},
100
  author = {Russell, Jenna and Glickenhaus, Ben and Thai, Katherine and Wieting, John and Iyyer, Mohit and Spero, Max and Emi, Bradley},
101
  journal = {arXiv preprint arXiv:2609.40295},
35.8m-g022-ai-r69.6/README.md CHANGED
@@ -18,7 +18,7 @@ tags:
18
 
19
  > **License: [CC BY-NC-SA 4.0](https://creativecommons.org/licenses/by-nc-sa/4.0/).** The WildAI code, models and dataset are released under the Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International license: you may use and adapt them for non-commercial purposes, with attribution, and must share adaptations under the same license.
20
 
21
- A 35.8M decoder-only language model trained from scratch for [*How Much Is an AI Token Worth? Scaling Laws for Wild AI-Generated Web Text*](https://arxiv.org/abs/2609.40295), one of 844 models that measure how wild AI-generated web text in pretraining data changes a model's loss. It is a research artifact for studying scaling laws, not an assistant: it has had no instruction tuning or safety training.
22
 
23
  ## Training data
24
 
@@ -95,7 +95,7 @@ The nanochat GPT recipe: rotary embeddings, RMSNorm without weights (also on que
95
  ## Citation
96
 
97
  ```bibtex
98
- @article{russell2026aitoken,
99
  title = {How Much Is an AI Token Worth? Scaling Laws for Wild AI-Generated Web Text},
100
  author = {Russell, Jenna and Glickenhaus, Ben and Thai, Katherine and Wieting, John and Iyyer, Mohit and Spero, Max and Emi, Bradley},
101
  journal = {arXiv preprint arXiv:2609.40295},
 
18
 
19
  > **License: [CC BY-NC-SA 4.0](https://creativecommons.org/licenses/by-nc-sa/4.0/).** The WildAI code, models and dataset are released under the Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International license: you may use and adapt them for non-commercial purposes, with attribution, and must share adaptations under the same license.
20
 
21
+ A 35.8M decoder-only language model trained from scratch for [*How Much Is an AI Token Worth? Scaling Laws for Wild AI-Generated Web Text*](https://arxiv.org/abs/2609.40295), one of the paper's models measuring how wild AI-generated web text in pretraining data changes a model's loss. It is a research artifact for studying scaling laws, not an assistant: it has had no instruction tuning or safety training.
22
 
23
  ## Training data
24
 
 
95
  ## Citation
96
 
97
  ```bibtex
98
+ @article{russell2026wildai,
99
  title = {How Much Is an AI Token Worth? Scaling Laws for Wild AI-Generated Web Text},
100
  author = {Russell, Jenna and Glickenhaus, Ben and Thai, Katherine and Wieting, John and Iyyer, Mohit and Spero, Max and Emi, Bradley},
101
  journal = {arXiv preprint arXiv:2609.40295},
35.8m-g022-ai-r8.76/README.md CHANGED
@@ -18,7 +18,7 @@ tags:
18
 
19
  > **License: [CC BY-NC-SA 4.0](https://creativecommons.org/licenses/by-nc-sa/4.0/).** The WildAI code, models and dataset are released under the Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International license: you may use and adapt them for non-commercial purposes, with attribution, and must share adaptations under the same license.
20
 
21
- A 35.8M decoder-only language model trained from scratch for [*How Much Is an AI Token Worth? Scaling Laws for Wild AI-Generated Web Text*](https://arxiv.org/abs/2609.40295), one of 844 models that measure how wild AI-generated web text in pretraining data changes a model's loss. It is a research artifact for studying scaling laws, not an assistant: it has had no instruction tuning or safety training.
22
 
23
  ## Training data
24
 
@@ -95,7 +95,7 @@ The nanochat GPT recipe: rotary embeddings, RMSNorm without weights (also on que
95
  ## Citation
96
 
97
  ```bibtex
98
- @article{russell2026aitoken,
99
  title = {How Much Is an AI Token Worth? Scaling Laws for Wild AI-Generated Web Text},
100
  author = {Russell, Jenna and Glickenhaus, Ben and Thai, Katherine and Wieting, John and Iyyer, Mohit and Spero, Max and Emi, Bradley},
101
  journal = {arXiv preprint arXiv:2609.40295},
 
18
 
19
  > **License: [CC BY-NC-SA 4.0](https://creativecommons.org/licenses/by-nc-sa/4.0/).** The WildAI code, models and dataset are released under the Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International license: you may use and adapt them for non-commercial purposes, with attribution, and must share adaptations under the same license.
20
 
21
+ A 35.8M decoder-only language model trained from scratch for [*How Much Is an AI Token Worth? Scaling Laws for Wild AI-Generated Web Text*](https://arxiv.org/abs/2609.40295), one of the paper's models measuring how wild AI-generated web text in pretraining data changes a model's loss. It is a research artifact for studying scaling laws, not an assistant: it has had no instruction tuning or safety training.
22
 
23
  ## Training data
24
 
 
95
  ## Citation
96
 
97
  ```bibtex
98
+ @article{russell2026wildai,
99
  title = {How Much Is an AI Token Worth? Scaling Laws for Wild AI-Generated Web Text},
100
  author = {Russell, Jenna and Glickenhaus, Ben and Thai, Katherine and Wieting, John and Iyyer, Mohit and Spero, Max and Emi, Bradley},
101
  journal = {arXiv preprint arXiv:2609.40295},
35.8m-g022-control/README.md CHANGED
@@ -18,7 +18,7 @@ tags:
18
 
19
  > **License: [CC BY-NC-SA 4.0](https://creativecommons.org/licenses/by-nc-sa/4.0/).** The WildAI code, models and dataset are released under the Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International license: you may use and adapt them for non-commercial purposes, with attribution, and must share adaptations under the same license.
20
 
21
- A 35.8M decoder-only language model trained from scratch for [*How Much Is an AI Token Worth? Scaling Laws for Wild AI-Generated Web Text*](https://arxiv.org/abs/2609.40295), one of 844 models that measure how wild AI-generated web text in pretraining data changes a model's loss. It is a research artifact for studying scaling laws, not an assistant: it has had no instruction tuning or safety training.
22
 
23
  ## Training data
24
 
@@ -95,7 +95,7 @@ The nanochat GPT recipe: rotary embeddings, RMSNorm without weights (also on que
95
  ## Citation
96
 
97
  ```bibtex
98
- @article{russell2026aitoken,
99
  title = {How Much Is an AI Token Worth? Scaling Laws for Wild AI-Generated Web Text},
100
  author = {Russell, Jenna and Glickenhaus, Ben and Thai, Katherine and Wieting, John and Iyyer, Mohit and Spero, Max and Emi, Bradley},
101
  journal = {arXiv preprint arXiv:2609.40295},
 
18
 
19
  > **License: [CC BY-NC-SA 4.0](https://creativecommons.org/licenses/by-nc-sa/4.0/).** The WildAI code, models and dataset are released under the Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International license: you may use and adapt them for non-commercial purposes, with attribution, and must share adaptations under the same license.
20
 
21
+ A 35.8M decoder-only language model trained from scratch for [*How Much Is an AI Token Worth? Scaling Laws for Wild AI-Generated Web Text*](https://arxiv.org/abs/2609.40295), one of the paper's models measuring how wild AI-generated web text in pretraining data changes a model's loss. It is a research artifact for studying scaling laws, not an assistant: it has had no instruction tuning or safety training.
22
 
23
  ## Training data
24
 
 
95
  ## Citation
96
 
97
  ```bibtex
98
+ @article{russell2026wildai,
99
  title = {How Much Is an AI Token Worth? Scaling Laws for Wild AI-Generated Web Text},
100
  author = {Russell, Jenna and Glickenhaus, Ben and Thai, Katherine and Wieting, John and Iyyer, Mohit and Spero, Max and Emi, Bradley},
101
  journal = {arXiv preprint arXiv:2609.40295},
35.8m-g022-human-r0.025/README.md CHANGED
@@ -18,7 +18,7 @@ tags:
18
 
19
  > **License: [CC BY-NC-SA 4.0](https://creativecommons.org/licenses/by-nc-sa/4.0/).** The WildAI code, models and dataset are released under the Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International license: you may use and adapt them for non-commercial purposes, with attribution, and must share adaptations under the same license.
20
 
21
- A 35.8M decoder-only language model trained from scratch for [*How Much Is an AI Token Worth? Scaling Laws for Wild AI-Generated Web Text*](https://arxiv.org/abs/2609.40295), one of 844 models that measure how wild AI-generated web text in pretraining data changes a model's loss. It is a research artifact for studying scaling laws, not an assistant: it has had no instruction tuning or safety training.
22
 
23
  ## Training data
24
 
@@ -95,7 +95,7 @@ The nanochat GPT recipe: rotary embeddings, RMSNorm without weights (also on que
95
  ## Citation
96
 
97
  ```bibtex
98
- @article{russell2026aitoken,
99
  title = {How Much Is an AI Token Worth? Scaling Laws for Wild AI-Generated Web Text},
100
  author = {Russell, Jenna and Glickenhaus, Ben and Thai, Katherine and Wieting, John and Iyyer, Mohit and Spero, Max and Emi, Bradley},
101
  journal = {arXiv preprint arXiv:2609.40295},
 
18
 
19
  > **License: [CC BY-NC-SA 4.0](https://creativecommons.org/licenses/by-nc-sa/4.0/).** The WildAI code, models and dataset are released under the Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International license: you may use and adapt them for non-commercial purposes, with attribution, and must share adaptations under the same license.
20
 
21
+ A 35.8M decoder-only language model trained from scratch for [*How Much Is an AI Token Worth? Scaling Laws for Wild AI-Generated Web Text*](https://arxiv.org/abs/2609.40295), one of the paper's models measuring how wild AI-generated web text in pretraining data changes a model's loss. It is a research artifact for studying scaling laws, not an assistant: it has had no instruction tuning or safety training.
22
 
23
  ## Training data
24
 
 
95
  ## Citation
96
 
97
  ```bibtex
98
+ @article{russell2026wildai,
99
  title = {How Much Is an AI Token Worth? Scaling Laws for Wild AI-Generated Web Text},
100
  author = {Russell, Jenna and Glickenhaus, Ben and Thai, Katherine and Wieting, John and Iyyer, Mohit and Spero, Max and Emi, Bradley},
101
  journal = {arXiv preprint arXiv:2609.40295},
35.8m-g022-human-r0.05/README.md CHANGED
@@ -18,7 +18,7 @@ tags:
18
 
19
  > **License: [CC BY-NC-SA 4.0](https://creativecommons.org/licenses/by-nc-sa/4.0/).** The WildAI code, models and dataset are released under the Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International license: you may use and adapt them for non-commercial purposes, with attribution, and must share adaptations under the same license.
20
 
21
- A 35.8M decoder-only language model trained from scratch for [*How Much Is an AI Token Worth? Scaling Laws for Wild AI-Generated Web Text*](https://arxiv.org/abs/2609.40295), one of 844 models that measure how wild AI-generated web text in pretraining data changes a model's loss. It is a research artifact for studying scaling laws, not an assistant: it has had no instruction tuning or safety training.
22
 
23
  ## Training data
24
 
@@ -95,7 +95,7 @@ The nanochat GPT recipe: rotary embeddings, RMSNorm without weights (also on que
95
  ## Citation
96
 
97
  ```bibtex
98
- @article{russell2026aitoken,
99
  title = {How Much Is an AI Token Worth? Scaling Laws for Wild AI-Generated Web Text},
100
  author = {Russell, Jenna and Glickenhaus, Ben and Thai, Katherine and Wieting, John and Iyyer, Mohit and Spero, Max and Emi, Bradley},
101
  journal = {arXiv preprint arXiv:2609.40295},
 
18
 
19
  > **License: [CC BY-NC-SA 4.0](https://creativecommons.org/licenses/by-nc-sa/4.0/).** The WildAI code, models and dataset are released under the Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International license: you may use and adapt them for non-commercial purposes, with attribution, and must share adaptations under the same license.
20
 
21
+ A 35.8M decoder-only language model trained from scratch for [*How Much Is an AI Token Worth? Scaling Laws for Wild AI-Generated Web Text*](https://arxiv.org/abs/2609.40295), one of the paper's models measuring how wild AI-generated web text in pretraining data changes a model's loss. It is a research artifact for studying scaling laws, not an assistant: it has had no instruction tuning or safety training.
22
 
23
  ## Training data
24
 
 
95
  ## Citation
96
 
97
  ```bibtex
98
+ @article{russell2026wildai,
99
  title = {How Much Is an AI Token Worth? Scaling Laws for Wild AI-Generated Web Text},
100
  author = {Russell, Jenna and Glickenhaus, Ben and Thai, Katherine and Wieting, John and Iyyer, Mohit and Spero, Max and Emi, Bradley},
101
  journal = {arXiv preprint arXiv:2609.40295},
35.8m-g022-human-r4/README.md CHANGED
@@ -18,7 +18,7 @@ tags:
18
 
19
  > **License: [CC BY-NC-SA 4.0](https://creativecommons.org/licenses/by-nc-sa/4.0/).** The WildAI code, models and dataset are released under the Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International license: you may use and adapt them for non-commercial purposes, with attribution, and must share adaptations under the same license.
20
 
21
- A 35.8M decoder-only language model trained from scratch for [*How Much Is an AI Token Worth? Scaling Laws for Wild AI-Generated Web Text*](https://arxiv.org/abs/2609.40295), one of 844 models that measure how wild AI-generated web text in pretraining data changes a model's loss. It is a research artifact for studying scaling laws, not an assistant: it has had no instruction tuning or safety training.
22
 
23
  ## Training data
24
 
@@ -95,7 +95,7 @@ The nanochat GPT recipe: rotary embeddings, RMSNorm without weights (also on que
95
  ## Citation
96
 
97
  ```bibtex
98
- @article{russell2026aitoken,
99
  title = {How Much Is an AI Token Worth? Scaling Laws for Wild AI-Generated Web Text},
100
  author = {Russell, Jenna and Glickenhaus, Ben and Thai, Katherine and Wieting, John and Iyyer, Mohit and Spero, Max and Emi, Bradley},
101
  journal = {arXiv preprint arXiv:2609.40295},
 
18
 
19
  > **License: [CC BY-NC-SA 4.0](https://creativecommons.org/licenses/by-nc-sa/4.0/).** The WildAI code, models and dataset are released under the Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International license: you may use and adapt them for non-commercial purposes, with attribution, and must share adaptations under the same license.
20
 
21
+ A 35.8M decoder-only language model trained from scratch for [*How Much Is an AI Token Worth? Scaling Laws for Wild AI-Generated Web Text*](https://arxiv.org/abs/2609.40295), one of the paper's models measuring how wild AI-generated web text in pretraining data changes a model's loss. It is a research artifact for studying scaling laws, not an assistant: it has had no instruction tuning or safety training.
22
 
23
  ## Training data
24
 
 
95
  ## Citation
96
 
97
  ```bibtex
98
+ @article{russell2026wildai,
99
  title = {How Much Is an AI Token Worth? Scaling Laws for Wild AI-Generated Web Text},
100
  author = {Russell, Jenna and Glickenhaus, Ben and Thai, Katherine and Wieting, John and Iyyer, Mohit and Spero, Max and Emi, Bradley},
101
  journal = {arXiv preprint arXiv:2609.40295},
35.8m-g023-ai-r0.025/README.md CHANGED
@@ -18,7 +18,7 @@ tags:
18
 
19
  > **License: [CC BY-NC-SA 4.0](https://creativecommons.org/licenses/by-nc-sa/4.0/).** The WildAI code, models and dataset are released under the Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International license: you may use and adapt them for non-commercial purposes, with attribution, and must share adaptations under the same license.
20
 
21
- A 35.8M decoder-only language model trained from scratch for [*How Much Is an AI Token Worth? Scaling Laws for Wild AI-Generated Web Text*](https://arxiv.org/abs/2609.40295), one of 844 models that measure how wild AI-generated web text in pretraining data changes a model's loss. It is a research artifact for studying scaling laws, not an assistant: it has had no instruction tuning or safety training.
22
 
23
  ## Training data
24
 
@@ -95,7 +95,7 @@ The nanochat GPT recipe: rotary embeddings, RMSNorm without weights (also on que
95
  ## Citation
96
 
97
  ```bibtex
98
- @article{russell2026aitoken,
99
  title = {How Much Is an AI Token Worth? Scaling Laws for Wild AI-Generated Web Text},
100
  author = {Russell, Jenna and Glickenhaus, Ben and Thai, Katherine and Wieting, John and Iyyer, Mohit and Spero, Max and Emi, Bradley},
101
  journal = {arXiv preprint arXiv:2609.40295},
 
18
 
19
  > **License: [CC BY-NC-SA 4.0](https://creativecommons.org/licenses/by-nc-sa/4.0/).** The WildAI code, models and dataset are released under the Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International license: you may use and adapt them for non-commercial purposes, with attribution, and must share adaptations under the same license.
20
 
21
+ A 35.8M decoder-only language model trained from scratch for [*How Much Is an AI Token Worth? Scaling Laws for Wild AI-Generated Web Text*](https://arxiv.org/abs/2609.40295), one of the paper's models measuring how wild AI-generated web text in pretraining data changes a model's loss. It is a research artifact for studying scaling laws, not an assistant: it has had no instruction tuning or safety training.
22
 
23
  ## Training data
24
 
 
95
  ## Citation
96
 
97
  ```bibtex
98
+ @article{russell2026wildai,
99
  title = {How Much Is an AI Token Worth? Scaling Laws for Wild AI-Generated Web Text},
100
  author = {Russell, Jenna and Glickenhaus, Ben and Thai, Katherine and Wieting, John and Iyyer, Mohit and Spero, Max and Emi, Bradley},
101
  journal = {arXiv preprint arXiv:2609.40295},
35.8m-g023-ai-r0.05/README.md CHANGED
@@ -18,7 +18,7 @@ tags:
18
 
19
  > **License: [CC BY-NC-SA 4.0](https://creativecommons.org/licenses/by-nc-sa/4.0/).** The WildAI code, models and dataset are released under the Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International license: you may use and adapt them for non-commercial purposes, with attribution, and must share adaptations under the same license.
20
 
21
- A 35.8M decoder-only language model trained from scratch for [*How Much Is an AI Token Worth? Scaling Laws for Wild AI-Generated Web Text*](https://arxiv.org/abs/2609.40295), one of 844 models that measure how wild AI-generated web text in pretraining data changes a model's loss. It is a research artifact for studying scaling laws, not an assistant: it has had no instruction tuning or safety training.
22
 
23
  ## Training data
24
 
@@ -95,7 +95,7 @@ The nanochat GPT recipe: rotary embeddings, RMSNorm without weights (also on que
95
  ## Citation
96
 
97
  ```bibtex
98
- @article{russell2026aitoken,
99
  title = {How Much Is an AI Token Worth? Scaling Laws for Wild AI-Generated Web Text},
100
  author = {Russell, Jenna and Glickenhaus, Ben and Thai, Katherine and Wieting, John and Iyyer, Mohit and Spero, Max and Emi, Bradley},
101
  journal = {arXiv preprint arXiv:2609.40295},
 
18
 
19
  > **License: [CC BY-NC-SA 4.0](https://creativecommons.org/licenses/by-nc-sa/4.0/).** The WildAI code, models and dataset are released under the Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International license: you may use and adapt them for non-commercial purposes, with attribution, and must share adaptations under the same license.
20
 
21
+ A 35.8M decoder-only language model trained from scratch for [*How Much Is an AI Token Worth? Scaling Laws for Wild AI-Generated Web Text*](https://arxiv.org/abs/2609.40295), one of the paper's models measuring how wild AI-generated web text in pretraining data changes a model's loss. It is a research artifact for studying scaling laws, not an assistant: it has had no instruction tuning or safety training.
22
 
23
  ## Training data
24
 
 
95
  ## Citation
96
 
97
  ```bibtex
98
+ @article{russell2026wildai,
99
  title = {How Much Is an AI Token Worth? Scaling Laws for Wild AI-Generated Web Text},
100
  author = {Russell, Jenna and Glickenhaus, Ben and Thai, Katherine and Wieting, John and Iyyer, Mohit and Spero, Max and Emi, Bradley},
101
  journal = {arXiv preprint arXiv:2609.40295},
35.8m-g023-ai-r0.1/README.md CHANGED
@@ -18,7 +18,7 @@ tags:
18
 
19
  > **License: [CC BY-NC-SA 4.0](https://creativecommons.org/licenses/by-nc-sa/4.0/).** The WildAI code, models and dataset are released under the Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International license: you may use and adapt them for non-commercial purposes, with attribution, and must share adaptations under the same license.
20
 
21
- A 35.8M decoder-only language model trained from scratch for [*How Much Is an AI Token Worth? Scaling Laws for Wild AI-Generated Web Text*](https://arxiv.org/abs/2609.40295), one of 844 models that measure how wild AI-generated web text in pretraining data changes a model's loss. It is a research artifact for studying scaling laws, not an assistant: it has had no instruction tuning or safety training.
22
 
23
  ## Training data
24
 
@@ -95,7 +95,7 @@ The nanochat GPT recipe: rotary embeddings, RMSNorm without weights (also on que
95
  ## Citation
96
 
97
  ```bibtex
98
- @article{russell2026aitoken,
99
  title = {How Much Is an AI Token Worth? Scaling Laws for Wild AI-Generated Web Text},
100
  author = {Russell, Jenna and Glickenhaus, Ben and Thai, Katherine and Wieting, John and Iyyer, Mohit and Spero, Max and Emi, Bradley},
101
  journal = {arXiv preprint arXiv:2609.40295},
 
18
 
19
  > **License: [CC BY-NC-SA 4.0](https://creativecommons.org/licenses/by-nc-sa/4.0/).** The WildAI code, models and dataset are released under the Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International license: you may use and adapt them for non-commercial purposes, with attribution, and must share adaptations under the same license.
20
 
21
+ A 35.8M decoder-only language model trained from scratch for [*How Much Is an AI Token Worth? Scaling Laws for Wild AI-Generated Web Text*](https://arxiv.org/abs/2609.40295), one of the paper's models measuring how wild AI-generated web text in pretraining data changes a model's loss. It is a research artifact for studying scaling laws, not an assistant: it has had no instruction tuning or safety training.
22
 
23
  ## Training data
24
 
 
95
  ## Citation
96
 
97
  ```bibtex
98
+ @article{russell2026wildai,
99
  title = {How Much Is an AI Token Worth? Scaling Laws for Wild AI-Generated Web Text},
100
  author = {Russell, Jenna and Glickenhaus, Ben and Thai, Katherine and Wieting, John and Iyyer, Mohit and Spero, Max and Emi, Bradley},
101
  journal = {arXiv preprint arXiv:2609.40295},
35.8m-g023-ai-r0.25/README.md CHANGED
@@ -18,7 +18,7 @@ tags:
18
 
19
  > **License: [CC BY-NC-SA 4.0](https://creativecommons.org/licenses/by-nc-sa/4.0/).** The WildAI code, models and dataset are released under the Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International license: you may use and adapt them for non-commercial purposes, with attribution, and must share adaptations under the same license.
20
 
21
- A 35.8M decoder-only language model trained from scratch for [*How Much Is an AI Token Worth? Scaling Laws for Wild AI-Generated Web Text*](https://arxiv.org/abs/2609.40295), one of 844 models that measure how wild AI-generated web text in pretraining data changes a model's loss. It is a research artifact for studying scaling laws, not an assistant: it has had no instruction tuning or safety training.
22
 
23
  ## Training data
24
 
@@ -95,7 +95,7 @@ The nanochat GPT recipe: rotary embeddings, RMSNorm without weights (also on que
95
  ## Citation
96
 
97
  ```bibtex
98
- @article{russell2026aitoken,
99
  title = {How Much Is an AI Token Worth? Scaling Laws for Wild AI-Generated Web Text},
100
  author = {Russell, Jenna and Glickenhaus, Ben and Thai, Katherine and Wieting, John and Iyyer, Mohit and Spero, Max and Emi, Bradley},
101
  journal = {arXiv preprint arXiv:2609.40295},
 
18
 
19
  > **License: [CC BY-NC-SA 4.0](https://creativecommons.org/licenses/by-nc-sa/4.0/).** The WildAI code, models and dataset are released under the Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International license: you may use and adapt them for non-commercial purposes, with attribution, and must share adaptations under the same license.
20
 
21
+ A 35.8M decoder-only language model trained from scratch for [*How Much Is an AI Token Worth? Scaling Laws for Wild AI-Generated Web Text*](https://arxiv.org/abs/2609.40295), one of the paper's models measuring how wild AI-generated web text in pretraining data changes a model's loss. It is a research artifact for studying scaling laws, not an assistant: it has had no instruction tuning or safety training.
22
 
23
  ## Training data
24
 
 
95
  ## Citation
96
 
97
  ```bibtex
98
+ @article{russell2026wildai,
99
  title = {How Much Is an AI Token Worth? Scaling Laws for Wild AI-Generated Web Text},
100
  author = {Russell, Jenna and Glickenhaus, Ben and Thai, Katherine and Wieting, John and Iyyer, Mohit and Spero, Max and Emi, Bradley},
101
  journal = {arXiv preprint arXiv:2609.40295},
35.8m-g023-ai-r0.5/README.md CHANGED
@@ -18,7 +18,7 @@ tags:
18
 
19
  > **License: [CC BY-NC-SA 4.0](https://creativecommons.org/licenses/by-nc-sa/4.0/).** The WildAI code, models and dataset are released under the Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International license: you may use and adapt them for non-commercial purposes, with attribution, and must share adaptations under the same license.
20
 
21
- A 35.8M decoder-only language model trained from scratch for [*How Much Is an AI Token Worth? Scaling Laws for Wild AI-Generated Web Text*](https://arxiv.org/abs/2609.40295), one of 844 models that measure how wild AI-generated web text in pretraining data changes a model's loss. It is a research artifact for studying scaling laws, not an assistant: it has had no instruction tuning or safety training.
22
 
23
  ## Training data
24
 
@@ -95,7 +95,7 @@ The nanochat GPT recipe: rotary embeddings, RMSNorm without weights (also on que
95
  ## Citation
96
 
97
  ```bibtex
98
- @article{russell2026aitoken,
99
  title = {How Much Is an AI Token Worth? Scaling Laws for Wild AI-Generated Web Text},
100
  author = {Russell, Jenna and Glickenhaus, Ben and Thai, Katherine and Wieting, John and Iyyer, Mohit and Spero, Max and Emi, Bradley},
101
  journal = {arXiv preprint arXiv:2609.40295},
 
18
 
19
  > **License: [CC BY-NC-SA 4.0](https://creativecommons.org/licenses/by-nc-sa/4.0/).** The WildAI code, models and dataset are released under the Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International license: you may use and adapt them for non-commercial purposes, with attribution, and must share adaptations under the same license.
20
 
21
+ A 35.8M decoder-only language model trained from scratch for [*How Much Is an AI Token Worth? Scaling Laws for Wild AI-Generated Web Text*](https://arxiv.org/abs/2609.40295), one of the paper's models measuring how wild AI-generated web text in pretraining data changes a model's loss. It is a research artifact for studying scaling laws, not an assistant: it has had no instruction tuning or safety training.
22
 
23
  ## Training data
24
 
 
95
  ## Citation
96
 
97
  ```bibtex
98
+ @article{russell2026wildai,
99
  title = {How Much Is an AI Token Worth? Scaling Laws for Wild AI-Generated Web Text},
100
  author = {Russell, Jenna and Glickenhaus, Ben and Thai, Katherine and Wieting, John and Iyyer, Mohit and Spero, Max and Emi, Bradley},
101
  journal = {arXiv preprint arXiv:2609.40295},
35.8m-g023-ai-r1/README.md CHANGED
@@ -18,7 +18,7 @@ tags:
18
 
19
  > **License: [CC BY-NC-SA 4.0](https://creativecommons.org/licenses/by-nc-sa/4.0/).** The WildAI code, models and dataset are released under the Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International license: you may use and adapt them for non-commercial purposes, with attribution, and must share adaptations under the same license.
20
 
21
- A 35.8M decoder-only language model trained from scratch for [*How Much Is an AI Token Worth? Scaling Laws for Wild AI-Generated Web Text*](https://arxiv.org/abs/2609.40295), one of 844 models that measure how wild AI-generated web text in pretraining data changes a model's loss. It is a research artifact for studying scaling laws, not an assistant: it has had no instruction tuning or safety training.
22
 
23
  ## Training data
24
 
@@ -95,7 +95,7 @@ The nanochat GPT recipe: rotary embeddings, RMSNorm without weights (also on que
95
  ## Citation
96
 
97
  ```bibtex
98
- @article{russell2026aitoken,
99
  title = {How Much Is an AI Token Worth? Scaling Laws for Wild AI-Generated Web Text},
100
  author = {Russell, Jenna and Glickenhaus, Ben and Thai, Katherine and Wieting, John and Iyyer, Mohit and Spero, Max and Emi, Bradley},
101
  journal = {arXiv preprint arXiv:2609.40295},
 
18
 
19
  > **License: [CC BY-NC-SA 4.0](https://creativecommons.org/licenses/by-nc-sa/4.0/).** The WildAI code, models and dataset are released under the Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International license: you may use and adapt them for non-commercial purposes, with attribution, and must share adaptations under the same license.
20
 
21
+ A 35.8M decoder-only language model trained from scratch for [*How Much Is an AI Token Worth? Scaling Laws for Wild AI-Generated Web Text*](https://arxiv.org/abs/2609.40295), one of the paper's models measuring how wild AI-generated web text in pretraining data changes a model's loss. It is a research artifact for studying scaling laws, not an assistant: it has had no instruction tuning or safety training.
22
 
23
  ## Training data
24
 
 
95
  ## Citation
96
 
97
  ```bibtex
98
+ @article{russell2026wildai,
99
  title = {How Much Is an AI Token Worth? Scaling Laws for Wild AI-Generated Web Text},
100
  author = {Russell, Jenna and Glickenhaus, Ben and Thai, Katherine and Wieting, John and Iyyer, Mohit and Spero, Max and Emi, Bradley},
101
  journal = {arXiv preprint arXiv:2609.40295},
35.8m-g023-ai-r16/README.md CHANGED
@@ -18,7 +18,7 @@ tags:
18
 
19
  > **License: [CC BY-NC-SA 4.0](https://creativecommons.org/licenses/by-nc-sa/4.0/).** The WildAI code, models and dataset are released under the Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International license: you may use and adapt them for non-commercial purposes, with attribution, and must share adaptations under the same license.
20
 
21
- A 35.8M decoder-only language model trained from scratch for [*How Much Is an AI Token Worth? Scaling Laws for Wild AI-Generated Web Text*](https://arxiv.org/abs/2609.40295), one of 844 models that measure how wild AI-generated web text in pretraining data changes a model's loss. It is a research artifact for studying scaling laws, not an assistant: it has had no instruction tuning or safety training.
22
 
23
  ## Training data
24
 
@@ -95,7 +95,7 @@ The nanochat GPT recipe: rotary embeddings, RMSNorm without weights (also on que
95
  ## Citation
96
 
97
  ```bibtex
98
- @article{russell2026aitoken,
99
  title = {How Much Is an AI Token Worth? Scaling Laws for Wild AI-Generated Web Text},
100
  author = {Russell, Jenna and Glickenhaus, Ben and Thai, Katherine and Wieting, John and Iyyer, Mohit and Spero, Max and Emi, Bradley},
101
  journal = {arXiv preprint arXiv:2609.40295},
 
18
 
19
  > **License: [CC BY-NC-SA 4.0](https://creativecommons.org/licenses/by-nc-sa/4.0/).** The WildAI code, models and dataset are released under the Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International license: you may use and adapt them for non-commercial purposes, with attribution, and must share adaptations under the same license.
20
 
21
+ A 35.8M decoder-only language model trained from scratch for [*How Much Is an AI Token Worth? Scaling Laws for Wild AI-Generated Web Text*](https://arxiv.org/abs/2609.40295), one of the paper's models measuring how wild AI-generated web text in pretraining data changes a model's loss. It is a research artifact for studying scaling laws, not an assistant: it has had no instruction tuning or safety training.
22
 
23
  ## Training data
24
 
 
95
  ## Citation
96
 
97
  ```bibtex
98
+ @article{russell2026wildai,
99
  title = {How Much Is an AI Token Worth? Scaling Laws for Wild AI-Generated Web Text},
100
  author = {Russell, Jenna and Glickenhaus, Ben and Thai, Katherine and Wieting, John and Iyyer, Mohit and Spero, Max and Emi, Bradley},
101
  journal = {arXiv preprint arXiv:2609.40295},
35.8m-g023-ai-r2/README.md CHANGED
@@ -18,7 +18,7 @@ tags:
18
 
19
  > **License: [CC BY-NC-SA 4.0](https://creativecommons.org/licenses/by-nc-sa/4.0/).** The WildAI code, models and dataset are released under the Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International license: you may use and adapt them for non-commercial purposes, with attribution, and must share adaptations under the same license.
20
 
21
- A 35.8M decoder-only language model trained from scratch for [*How Much Is an AI Token Worth? Scaling Laws for Wild AI-Generated Web Text*](https://arxiv.org/abs/2609.40295), one of 844 models that measure how wild AI-generated web text in pretraining data changes a model's loss. It is a research artifact for studying scaling laws, not an assistant: it has had no instruction tuning or safety training.
22
 
23
  ## Training data
24
 
@@ -95,7 +95,7 @@ The nanochat GPT recipe: rotary embeddings, RMSNorm without weights (also on que
95
  ## Citation
96
 
97
  ```bibtex
98
- @article{russell2026aitoken,
99
  title = {How Much Is an AI Token Worth? Scaling Laws for Wild AI-Generated Web Text},
100
  author = {Russell, Jenna and Glickenhaus, Ben and Thai, Katherine and Wieting, John and Iyyer, Mohit and Spero, Max and Emi, Bradley},
101
  journal = {arXiv preprint arXiv:2609.40295},
 
18
 
19
  > **License: [CC BY-NC-SA 4.0](https://creativecommons.org/licenses/by-nc-sa/4.0/).** The WildAI code, models and dataset are released under the Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International license: you may use and adapt them for non-commercial purposes, with attribution, and must share adaptations under the same license.
20
 
21
+ A 35.8M decoder-only language model trained from scratch for [*How Much Is an AI Token Worth? Scaling Laws for Wild AI-Generated Web Text*](https://arxiv.org/abs/2609.40295), one of the paper's models measuring how wild AI-generated web text in pretraining data changes a model's loss. It is a research artifact for studying scaling laws, not an assistant: it has had no instruction tuning or safety training.
22
 
23
  ## Training data
24
 
 
95
  ## Citation
96
 
97
  ```bibtex
98
+ @article{russell2026wildai,
99
  title = {How Much Is an AI Token Worth? Scaling Laws for Wild AI-Generated Web Text},
100
  author = {Russell, Jenna and Glickenhaus, Ben and Thai, Katherine and Wieting, John and Iyyer, Mohit and Spero, Max and Emi, Bradley},
101
  journal = {arXiv preprint arXiv:2609.40295},
35.8m-g023-ai-r32/README.md CHANGED
@@ -18,7 +18,7 @@ tags:
18
 
19
  > **License: [CC BY-NC-SA 4.0](https://creativecommons.org/licenses/by-nc-sa/4.0/).** The WildAI code, models and dataset are released under the Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International license: you may use and adapt them for non-commercial purposes, with attribution, and must share adaptations under the same license.
20
 
21
- A 35.8M decoder-only language model trained from scratch for [*How Much Is an AI Token Worth? Scaling Laws for Wild AI-Generated Web Text*](https://arxiv.org/abs/2609.40295), one of 844 models that measure how wild AI-generated web text in pretraining data changes a model's loss. It is a research artifact for studying scaling laws, not an assistant: it has had no instruction tuning or safety training.
22
 
23
  ## Training data
24
 
@@ -95,7 +95,7 @@ The nanochat GPT recipe: rotary embeddings, RMSNorm without weights (also on que
95
  ## Citation
96
 
97
  ```bibtex
98
- @article{russell2026aitoken,
99
  title = {How Much Is an AI Token Worth? Scaling Laws for Wild AI-Generated Web Text},
100
  author = {Russell, Jenna and Glickenhaus, Ben and Thai, Katherine and Wieting, John and Iyyer, Mohit and Spero, Max and Emi, Bradley},
101
  journal = {arXiv preprint arXiv:2609.40295},
 
18
 
19
  > **License: [CC BY-NC-SA 4.0](https://creativecommons.org/licenses/by-nc-sa/4.0/).** The WildAI code, models and dataset are released under the Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International license: you may use and adapt them for non-commercial purposes, with attribution, and must share adaptations under the same license.
20
 
21
+ A 35.8M decoder-only language model trained from scratch for [*How Much Is an AI Token Worth? Scaling Laws for Wild AI-Generated Web Text*](https://arxiv.org/abs/2609.40295), one of the paper's models measuring how wild AI-generated web text in pretraining data changes a model's loss. It is a research artifact for studying scaling laws, not an assistant: it has had no instruction tuning or safety training.
22
 
23
  ## Training data
24
 
 
95
  ## Citation
96
 
97
  ```bibtex
98
+ @article{russell2026wildai,
99
  title = {How Much Is an AI Token Worth? Scaling Laws for Wild AI-Generated Web Text},
100
  author = {Russell, Jenna and Glickenhaus, Ben and Thai, Katherine and Wieting, John and Iyyer, Mohit and Spero, Max and Emi, Bradley},
101
  journal = {arXiv preprint arXiv:2609.40295},
35.8m-g023-ai-r4/README.md CHANGED
@@ -18,7 +18,7 @@ tags:
18
 
19
  > **License: [CC BY-NC-SA 4.0](https://creativecommons.org/licenses/by-nc-sa/4.0/).** The WildAI code, models and dataset are released under the Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International license: you may use and adapt them for non-commercial purposes, with attribution, and must share adaptations under the same license.
20
 
21
- A 35.8M decoder-only language model trained from scratch for [*How Much Is an AI Token Worth? Scaling Laws for Wild AI-Generated Web Text*](https://arxiv.org/abs/2609.40295), one of 844 models that measure how wild AI-generated web text in pretraining data changes a model's loss. It is a research artifact for studying scaling laws, not an assistant: it has had no instruction tuning or safety training.
22
 
23
  ## Training data
24
 
@@ -95,7 +95,7 @@ The nanochat GPT recipe: rotary embeddings, RMSNorm without weights (also on que
95
  ## Citation
96
 
97
  ```bibtex
98
- @article{russell2026aitoken,
99
  title = {How Much Is an AI Token Worth? Scaling Laws for Wild AI-Generated Web Text},
100
  author = {Russell, Jenna and Glickenhaus, Ben and Thai, Katherine and Wieting, John and Iyyer, Mohit and Spero, Max and Emi, Bradley},
101
  journal = {arXiv preprint arXiv:2609.40295},
 
18
 
19
  > **License: [CC BY-NC-SA 4.0](https://creativecommons.org/licenses/by-nc-sa/4.0/).** The WildAI code, models and dataset are released under the Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International license: you may use and adapt them for non-commercial purposes, with attribution, and must share adaptations under the same license.
20
 
21
+ A 35.8M decoder-only language model trained from scratch for [*How Much Is an AI Token Worth? Scaling Laws for Wild AI-Generated Web Text*](https://arxiv.org/abs/2609.40295), one of the paper's models measuring how wild AI-generated web text in pretraining data changes a model's loss. It is a research artifact for studying scaling laws, not an assistant: it has had no instruction tuning or safety training.
22
 
23
  ## Training data
24
 
 
95
  ## Citation
96
 
97
  ```bibtex
98
+ @article{russell2026wildai,
99
  title = {How Much Is an AI Token Worth? Scaling Laws for Wild AI-Generated Web Text},
100
  author = {Russell, Jenna and Glickenhaus, Ben and Thai, Katherine and Wieting, John and Iyyer, Mohit and Spero, Max and Emi, Bradley},
101
  journal = {arXiv preprint arXiv:2609.40295},
35.8m-g023-ai-r64/README.md CHANGED
@@ -18,7 +18,7 @@ tags:
18
 
19
  > **License: [CC BY-NC-SA 4.0](https://creativecommons.org/licenses/by-nc-sa/4.0/).** The WildAI code, models and dataset are released under the Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International license: you may use and adapt them for non-commercial purposes, with attribution, and must share adaptations under the same license.
20
 
21
- A 35.8M decoder-only language model trained from scratch for [*How Much Is an AI Token Worth? Scaling Laws for Wild AI-Generated Web Text*](https://arxiv.org/abs/2609.40295), one of 844 models that measure how wild AI-generated web text in pretraining data changes a model's loss. It is a research artifact for studying scaling laws, not an assistant: it has had no instruction tuning or safety training.
22
 
23
  ## Training data
24
 
@@ -95,7 +95,7 @@ The nanochat GPT recipe: rotary embeddings, RMSNorm without weights (also on que
95
  ## Citation
96
 
97
  ```bibtex
98
- @article{russell2026aitoken,
99
  title = {How Much Is an AI Token Worth? Scaling Laws for Wild AI-Generated Web Text},
100
  author = {Russell, Jenna and Glickenhaus, Ben and Thai, Katherine and Wieting, John and Iyyer, Mohit and Spero, Max and Emi, Bradley},
101
  journal = {arXiv preprint arXiv:2609.40295},
 
18
 
19
  > **License: [CC BY-NC-SA 4.0](https://creativecommons.org/licenses/by-nc-sa/4.0/).** The WildAI code, models and dataset are released under the Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International license: you may use and adapt them for non-commercial purposes, with attribution, and must share adaptations under the same license.
20
 
21
+ A 35.8M decoder-only language model trained from scratch for [*How Much Is an AI Token Worth? Scaling Laws for Wild AI-Generated Web Text*](https://arxiv.org/abs/2609.40295), one of the paper's models measuring how wild AI-generated web text in pretraining data changes a model's loss. It is a research artifact for studying scaling laws, not an assistant: it has had no instruction tuning or safety training.
22
 
23
  ## Training data
24
 
 
95
  ## Citation
96
 
97
  ```bibtex
98
+ @article{russell2026wildai,
99
  title = {How Much Is an AI Token Worth? Scaling Laws for Wild AI-Generated Web Text},
100
  author = {Russell, Jenna and Glickenhaus, Ben and Thai, Katherine and Wieting, John and Iyyer, Mohit and Spero, Max and Emi, Bradley},
101
  journal = {arXiv preprint arXiv:2609.40295},
35.8m-g023-ai-r8/README.md CHANGED
@@ -18,7 +18,7 @@ tags:
18
 
19
  > **License: [CC BY-NC-SA 4.0](https://creativecommons.org/licenses/by-nc-sa/4.0/).** The WildAI code, models and dataset are released under the Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International license: you may use and adapt them for non-commercial purposes, with attribution, and must share adaptations under the same license.
20
 
21
- A 35.8M decoder-only language model trained from scratch for [*How Much Is an AI Token Worth? Scaling Laws for Wild AI-Generated Web Text*](https://arxiv.org/abs/2609.40295), one of 844 models that measure how wild AI-generated web text in pretraining data changes a model's loss. It is a research artifact for studying scaling laws, not an assistant: it has had no instruction tuning or safety training.
22
 
23
  ## Training data
24
 
@@ -95,7 +95,7 @@ The nanochat GPT recipe: rotary embeddings, RMSNorm without weights (also on que
95
  ## Citation
96
 
97
  ```bibtex
98
- @article{russell2026aitoken,
99
  title = {How Much Is an AI Token Worth? Scaling Laws for Wild AI-Generated Web Text},
100
  author = {Russell, Jenna and Glickenhaus, Ben and Thai, Katherine and Wieting, John and Iyyer, Mohit and Spero, Max and Emi, Bradley},
101
  journal = {arXiv preprint arXiv:2609.40295},
 
18
 
19
  > **License: [CC BY-NC-SA 4.0](https://creativecommons.org/licenses/by-nc-sa/4.0/).** The WildAI code, models and dataset are released under the Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International license: you may use and adapt them for non-commercial purposes, with attribution, and must share adaptations under the same license.
20
 
21
+ A 35.8M decoder-only language model trained from scratch for [*How Much Is an AI Token Worth? Scaling Laws for Wild AI-Generated Web Text*](https://arxiv.org/abs/2609.40295), one of the paper's models measuring how wild AI-generated web text in pretraining data changes a model's loss. It is a research artifact for studying scaling laws, not an assistant: it has had no instruction tuning or safety training.
22
 
23
  ## Training data
24
 
 
95
  ## Citation
96
 
97
  ```bibtex
98
+ @article{russell2026wildai,
99
  title = {How Much Is an AI Token Worth? Scaling Laws for Wild AI-Generated Web Text},
100
  author = {Russell, Jenna and Glickenhaus, Ben and Thai, Katherine and Wieting, John and Iyyer, Mohit and Spero, Max and Emi, Bradley},
101
  journal = {arXiv preprint arXiv:2609.40295},
35.8m-g023-control/README.md CHANGED
@@ -18,7 +18,7 @@ tags:
18
 
19
  > **License: [CC BY-NC-SA 4.0](https://creativecommons.org/licenses/by-nc-sa/4.0/).** The WildAI code, models and dataset are released under the Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International license: you may use and adapt them for non-commercial purposes, with attribution, and must share adaptations under the same license.
20
 
21
- A 35.8M decoder-only language model trained from scratch for [*How Much Is an AI Token Worth? Scaling Laws for Wild AI-Generated Web Text*](https://arxiv.org/abs/2609.40295), one of 844 models that measure how wild AI-generated web text in pretraining data changes a model's loss. It is a research artifact for studying scaling laws, not an assistant: it has had no instruction tuning or safety training.
22
 
23
  ## Training data
24
 
@@ -95,7 +95,7 @@ The nanochat GPT recipe: rotary embeddings, RMSNorm without weights (also on que
95
  ## Citation
96
 
97
  ```bibtex
98
- @article{russell2026aitoken,
99
  title = {How Much Is an AI Token Worth? Scaling Laws for Wild AI-Generated Web Text},
100
  author = {Russell, Jenna and Glickenhaus, Ben and Thai, Katherine and Wieting, John and Iyyer, Mohit and Spero, Max and Emi, Bradley},
101
  journal = {arXiv preprint arXiv:2609.40295},
 
18
 
19
  > **License: [CC BY-NC-SA 4.0](https://creativecommons.org/licenses/by-nc-sa/4.0/).** The WildAI code, models and dataset are released under the Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International license: you may use and adapt them for non-commercial purposes, with attribution, and must share adaptations under the same license.
20
 
21
+ A 35.8M decoder-only language model trained from scratch for [*How Much Is an AI Token Worth? Scaling Laws for Wild AI-Generated Web Text*](https://arxiv.org/abs/2609.40295), one of the paper's models measuring how wild AI-generated web text in pretraining data changes a model's loss. It is a research artifact for studying scaling laws, not an assistant: it has had no instruction tuning or safety training.
22
 
23
  ## Training data
24
 
 
95
  ## Citation
96
 
97
  ```bibtex
98
+ @article{russell2026wildai,
99
  title = {How Much Is an AI Token Worth? Scaling Laws for Wild AI-Generated Web Text},
100
  author = {Russell, Jenna and Glickenhaus, Ben and Thai, Katherine and Wieting, John and Iyyer, Mohit and Spero, Max and Emi, Bradley},
101
  journal = {arXiv preprint arXiv:2609.40295},
35.8m-g023-human-r0.025/README.md CHANGED
@@ -18,7 +18,7 @@ tags:
18
 
19
  > **License: [CC BY-NC-SA 4.0](https://creativecommons.org/licenses/by-nc-sa/4.0/).** The WildAI code, models and dataset are released under the Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International license: you may use and adapt them for non-commercial purposes, with attribution, and must share adaptations under the same license.
20
 
21
- A 35.8M decoder-only language model trained from scratch for [*How Much Is an AI Token Worth? Scaling Laws for Wild AI-Generated Web Text*](https://arxiv.org/abs/2609.40295), one of 844 models that measure how wild AI-generated web text in pretraining data changes a model's loss. It is a research artifact for studying scaling laws, not an assistant: it has had no instruction tuning or safety training.
22
 
23
  ## Training data
24
 
@@ -95,7 +95,7 @@ The nanochat GPT recipe: rotary embeddings, RMSNorm without weights (also on que
95
  ## Citation
96
 
97
  ```bibtex
98
- @article{russell2026aitoken,
99
  title = {How Much Is an AI Token Worth? Scaling Laws for Wild AI-Generated Web Text},
100
  author = {Russell, Jenna and Glickenhaus, Ben and Thai, Katherine and Wieting, John and Iyyer, Mohit and Spero, Max and Emi, Bradley},
101
  journal = {arXiv preprint arXiv:2609.40295},
 
18
 
19
  > **License: [CC BY-NC-SA 4.0](https://creativecommons.org/licenses/by-nc-sa/4.0/).** The WildAI code, models and dataset are released under the Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International license: you may use and adapt them for non-commercial purposes, with attribution, and must share adaptations under the same license.
20
 
21
+ A 35.8M decoder-only language model trained from scratch for [*How Much Is an AI Token Worth? Scaling Laws for Wild AI-Generated Web Text*](https://arxiv.org/abs/2609.40295), one of the paper's models measuring how wild AI-generated web text in pretraining data changes a model's loss. It is a research artifact for studying scaling laws, not an assistant: it has had no instruction tuning or safety training.
22
 
23
  ## Training data
24
 
 
95
  ## Citation
96
 
97
  ```bibtex
98
+ @article{russell2026wildai,
99
  title = {How Much Is an AI Token Worth? Scaling Laws for Wild AI-Generated Web Text},
100
  author = {Russell, Jenna and Glickenhaus, Ben and Thai, Katherine and Wieting, John and Iyyer, Mohit and Spero, Max and Emi, Bradley},
101
  journal = {arXiv preprint arXiv:2609.40295},
35.8m-g023-human-r0.05/README.md CHANGED
@@ -18,7 +18,7 @@ tags:
18
 
19
  > **License: [CC BY-NC-SA 4.0](https://creativecommons.org/licenses/by-nc-sa/4.0/).** The WildAI code, models and dataset are released under the Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International license: you may use and adapt them for non-commercial purposes, with attribution, and must share adaptations under the same license.
20
 
21
- A 35.8M decoder-only language model trained from scratch for [*How Much Is an AI Token Worth? Scaling Laws for Wild AI-Generated Web Text*](https://arxiv.org/abs/2609.40295), one of 844 models that measure how wild AI-generated web text in pretraining data changes a model's loss. It is a research artifact for studying scaling laws, not an assistant: it has had no instruction tuning or safety training.
22
 
23
  ## Training data
24
 
@@ -95,7 +95,7 @@ The nanochat GPT recipe: rotary embeddings, RMSNorm without weights (also on que
95
  ## Citation
96
 
97
  ```bibtex
98
- @article{russell2026aitoken,
99
  title = {How Much Is an AI Token Worth? Scaling Laws for Wild AI-Generated Web Text},
100
  author = {Russell, Jenna and Glickenhaus, Ben and Thai, Katherine and Wieting, John and Iyyer, Mohit and Spero, Max and Emi, Bradley},
101
  journal = {arXiv preprint arXiv:2609.40295},
 
18
 
19
  > **License: [CC BY-NC-SA 4.0](https://creativecommons.org/licenses/by-nc-sa/4.0/).** The WildAI code, models and dataset are released under the Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International license: you may use and adapt them for non-commercial purposes, with attribution, and must share adaptations under the same license.
20
 
21
+ A 35.8M decoder-only language model trained from scratch for [*How Much Is an AI Token Worth? Scaling Laws for Wild AI-Generated Web Text*](https://arxiv.org/abs/2609.40295), one of the paper's models measuring how wild AI-generated web text in pretraining data changes a model's loss. It is a research artifact for studying scaling laws, not an assistant: it has had no instruction tuning or safety training.
22
 
23
  ## Training data
24
 
 
95
  ## Citation
96
 
97
  ```bibtex
98
+ @article{russell2026wildai,
99
  title = {How Much Is an AI Token Worth? Scaling Laws for Wild AI-Generated Web Text},
100
  author = {Russell, Jenna and Glickenhaus, Ben and Thai, Katherine and Wieting, John and Iyyer, Mohit and Spero, Max and Emi, Bradley},
101
  journal = {arXiv preprint arXiv:2609.40295},
35.8m-g023-human-r4/README.md CHANGED
@@ -18,7 +18,7 @@ tags:
18
 
19
  > **License: [CC BY-NC-SA 4.0](https://creativecommons.org/licenses/by-nc-sa/4.0/).** The WildAI code, models and dataset are released under the Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International license: you may use and adapt them for non-commercial purposes, with attribution, and must share adaptations under the same license.
20
 
21
- A 35.8M decoder-only language model trained from scratch for [*How Much Is an AI Token Worth? Scaling Laws for Wild AI-Generated Web Text*](https://arxiv.org/abs/2609.40295), one of 844 models that measure how wild AI-generated web text in pretraining data changes a model's loss. It is a research artifact for studying scaling laws, not an assistant: it has had no instruction tuning or safety training.
22
 
23
  ## Training data
24
 
@@ -95,7 +95,7 @@ The nanochat GPT recipe: rotary embeddings, RMSNorm without weights (also on que
95
  ## Citation
96
 
97
  ```bibtex
98
- @article{russell2026aitoken,
99
  title = {How Much Is an AI Token Worth? Scaling Laws for Wild AI-Generated Web Text},
100
  author = {Russell, Jenna and Glickenhaus, Ben and Thai, Katherine and Wieting, John and Iyyer, Mohit and Spero, Max and Emi, Bradley},
101
  journal = {arXiv preprint arXiv:2609.40295},
 
18
 
19
  > **License: [CC BY-NC-SA 4.0](https://creativecommons.org/licenses/by-nc-sa/4.0/).** The WildAI code, models and dataset are released under the Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International license: you may use and adapt them for non-commercial purposes, with attribution, and must share adaptations under the same license.
20
 
21
+ A 35.8M decoder-only language model trained from scratch for [*How Much Is an AI Token Worth? Scaling Laws for Wild AI-Generated Web Text*](https://arxiv.org/abs/2609.40295), one of the paper's models measuring how wild AI-generated web text in pretraining data changes a model's loss. It is a research artifact for studying scaling laws, not an assistant: it has had no instruction tuning or safety training.
22
 
23
  ## Training data
24
 
 
95
  ## Citation
96
 
97
  ```bibtex
98
+ @article{russell2026wildai,
99
  title = {How Much Is an AI Token Worth? Scaling Laws for Wild AI-Generated Web Text},
100
  author = {Russell, Jenna and Glickenhaus, Ben and Thai, Katherine and Wieting, John and Iyyer, Mohit and Spero, Max and Emi, Bradley},
101
  journal = {arXiv preprint arXiv:2609.40295},
35.8m-g024-ai-r0.05/README.md CHANGED
@@ -18,7 +18,7 @@ tags:
18
 
19
  > **License: [CC BY-NC-SA 4.0](https://creativecommons.org/licenses/by-nc-sa/4.0/).** The WildAI code, models and dataset are released under the Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International license: you may use and adapt them for non-commercial purposes, with attribution, and must share adaptations under the same license.
20
 
21
- A 35.8M decoder-only language model trained from scratch for [*How Much Is an AI Token Worth? Scaling Laws for Wild AI-Generated Web Text*](https://arxiv.org/abs/2609.40295), one of 844 models that measure how wild AI-generated web text in pretraining data changes a model's loss. It is a research artifact for studying scaling laws, not an assistant: it has had no instruction tuning or safety training.
22
 
23
  ## Training data
24
 
@@ -95,7 +95,7 @@ The nanochat GPT recipe: rotary embeddings, RMSNorm without weights (also on que
95
  ## Citation
96
 
97
  ```bibtex
98
- @article{russell2026aitoken,
99
  title = {How Much Is an AI Token Worth? Scaling Laws for Wild AI-Generated Web Text},
100
  author = {Russell, Jenna and Glickenhaus, Ben and Thai, Katherine and Wieting, John and Iyyer, Mohit and Spero, Max and Emi, Bradley},
101
  journal = {arXiv preprint arXiv:2609.40295},
 
18
 
19
  > **License: [CC BY-NC-SA 4.0](https://creativecommons.org/licenses/by-nc-sa/4.0/).** The WildAI code, models and dataset are released under the Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International license: you may use and adapt them for non-commercial purposes, with attribution, and must share adaptations under the same license.
20
 
21
+ A 35.8M decoder-only language model trained from scratch for [*How Much Is an AI Token Worth? Scaling Laws for Wild AI-Generated Web Text*](https://arxiv.org/abs/2609.40295), one of the paper's models measuring how wild AI-generated web text in pretraining data changes a model's loss. It is a research artifact for studying scaling laws, not an assistant: it has had no instruction tuning or safety training.
22
 
23
  ## Training data
24
 
 
95
  ## Citation
96
 
97
  ```bibtex
98
+ @article{russell2026wildai,
99
  title = {How Much Is an AI Token Worth? Scaling Laws for Wild AI-Generated Web Text},
100
  author = {Russell, Jenna and Glickenhaus, Ben and Thai, Katherine and Wieting, John and Iyyer, Mohit and Spero, Max and Emi, Bradley},
101
  journal = {arXiv preprint arXiv:2609.40295},
35.8m-g024-ai-r0.1/README.md CHANGED
@@ -18,7 +18,7 @@ tags:
18
 
19
  > **License: [CC BY-NC-SA 4.0](https://creativecommons.org/licenses/by-nc-sa/4.0/).** The WildAI code, models and dataset are released under the Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International license: you may use and adapt them for non-commercial purposes, with attribution, and must share adaptations under the same license.
20
 
21
- A 35.8M decoder-only language model trained from scratch for [*How Much Is an AI Token Worth? Scaling Laws for Wild AI-Generated Web Text*](https://arxiv.org/abs/2609.40295), one of 844 models that measure how wild AI-generated web text in pretraining data changes a model's loss. It is a research artifact for studying scaling laws, not an assistant: it has had no instruction tuning or safety training.
22
 
23
  ## Training data
24
 
@@ -95,7 +95,7 @@ The nanochat GPT recipe: rotary embeddings, RMSNorm without weights (also on que
95
  ## Citation
96
 
97
  ```bibtex
98
- @article{russell2026aitoken,
99
  title = {How Much Is an AI Token Worth? Scaling Laws for Wild AI-Generated Web Text},
100
  author = {Russell, Jenna and Glickenhaus, Ben and Thai, Katherine and Wieting, John and Iyyer, Mohit and Spero, Max and Emi, Bradley},
101
  journal = {arXiv preprint arXiv:2609.40295},
 
18
 
19
  > **License: [CC BY-NC-SA 4.0](https://creativecommons.org/licenses/by-nc-sa/4.0/).** The WildAI code, models and dataset are released under the Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International license: you may use and adapt them for non-commercial purposes, with attribution, and must share adaptations under the same license.
20
 
21
+ A 35.8M decoder-only language model trained from scratch for [*How Much Is an AI Token Worth? Scaling Laws for Wild AI-Generated Web Text*](https://arxiv.org/abs/2609.40295), one of the paper's models measuring how wild AI-generated web text in pretraining data changes a model's loss. It is a research artifact for studying scaling laws, not an assistant: it has had no instruction tuning or safety training.
22
 
23
  ## Training data
24
 
 
95
  ## Citation
96
 
97
  ```bibtex
98
+ @article{russell2026wildai,
99
  title = {How Much Is an AI Token Worth? Scaling Laws for Wild AI-Generated Web Text},
100
  author = {Russell, Jenna and Glickenhaus, Ben and Thai, Katherine and Wieting, John and Iyyer, Mohit and Spero, Max and Emi, Bradley},
101
  journal = {arXiv preprint arXiv:2609.40295},
35.8m-g024-ai-r0.25/README.md CHANGED
@@ -18,7 +18,7 @@ tags:
18
 
19
  > **License: [CC BY-NC-SA 4.0](https://creativecommons.org/licenses/by-nc-sa/4.0/).** The WildAI code, models and dataset are released under the Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International license: you may use and adapt them for non-commercial purposes, with attribution, and must share adaptations under the same license.
20
 
21
- A 35.8M decoder-only language model trained from scratch for [*How Much Is an AI Token Worth? Scaling Laws for Wild AI-Generated Web Text*](https://arxiv.org/abs/2609.40295), one of 844 models that measure how wild AI-generated web text in pretraining data changes a model's loss. It is a research artifact for studying scaling laws, not an assistant: it has had no instruction tuning or safety training.
22
 
23
  ## Training data
24
 
@@ -95,7 +95,7 @@ The nanochat GPT recipe: rotary embeddings, RMSNorm without weights (also on que
95
  ## Citation
96
 
97
  ```bibtex
98
- @article{russell2026aitoken,
99
  title = {How Much Is an AI Token Worth? Scaling Laws for Wild AI-Generated Web Text},
100
  author = {Russell, Jenna and Glickenhaus, Ben and Thai, Katherine and Wieting, John and Iyyer, Mohit and Spero, Max and Emi, Bradley},
101
  journal = {arXiv preprint arXiv:2609.40295},
 
18
 
19
  > **License: [CC BY-NC-SA 4.0](https://creativecommons.org/licenses/by-nc-sa/4.0/).** The WildAI code, models and dataset are released under the Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International license: you may use and adapt them for non-commercial purposes, with attribution, and must share adaptations under the same license.
20
 
21
+ A 35.8M decoder-only language model trained from scratch for [*How Much Is an AI Token Worth? Scaling Laws for Wild AI-Generated Web Text*](https://arxiv.org/abs/2609.40295), one of the paper's models measuring how wild AI-generated web text in pretraining data changes a model's loss. It is a research artifact for studying scaling laws, not an assistant: it has had no instruction tuning or safety training.
22
 
23
  ## Training data
24
 
 
95
  ## Citation
96
 
97
  ```bibtex
98
+ @article{russell2026wildai,
99
  title = {How Much Is an AI Token Worth? Scaling Laws for Wild AI-Generated Web Text},
100
  author = {Russell, Jenna and Glickenhaus, Ben and Thai, Katherine and Wieting, John and Iyyer, Mohit and Spero, Max and Emi, Bradley},
101
  journal = {arXiv preprint arXiv:2609.40295},
35.8m-g024-ai-r0.5/README.md CHANGED
@@ -18,7 +18,7 @@ tags:
18
 
19
  > **License: [CC BY-NC-SA 4.0](https://creativecommons.org/licenses/by-nc-sa/4.0/).** The WildAI code, models and dataset are released under the Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International license: you may use and adapt them for non-commercial purposes, with attribution, and must share adaptations under the same license.
20
 
21
- A 35.8M decoder-only language model trained from scratch for [*How Much Is an AI Token Worth? Scaling Laws for Wild AI-Generated Web Text*](https://arxiv.org/abs/2609.40295), one of 844 models that measure how wild AI-generated web text in pretraining data changes a model's loss. It is a research artifact for studying scaling laws, not an assistant: it has had no instruction tuning or safety training.
22
 
23
  ## Training data
24
 
@@ -95,7 +95,7 @@ The nanochat GPT recipe: rotary embeddings, RMSNorm without weights (also on que
95
  ## Citation
96
 
97
  ```bibtex
98
- @article{russell2026aitoken,
99
  title = {How Much Is an AI Token Worth? Scaling Laws for Wild AI-Generated Web Text},
100
  author = {Russell, Jenna and Glickenhaus, Ben and Thai, Katherine and Wieting, John and Iyyer, Mohit and Spero, Max and Emi, Bradley},
101
  journal = {arXiv preprint arXiv:2609.40295},
 
18
 
19
  > **License: [CC BY-NC-SA 4.0](https://creativecommons.org/licenses/by-nc-sa/4.0/).** The WildAI code, models and dataset are released under the Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International license: you may use and adapt them for non-commercial purposes, with attribution, and must share adaptations under the same license.
20
 
21
+ A 35.8M decoder-only language model trained from scratch for [*How Much Is an AI Token Worth? Scaling Laws for Wild AI-Generated Web Text*](https://arxiv.org/abs/2609.40295), one of the paper's models measuring how wild AI-generated web text in pretraining data changes a model's loss. It is a research artifact for studying scaling laws, not an assistant: it has had no instruction tuning or safety training.
22
 
23
  ## Training data
24
 
 
95
  ## Citation
96
 
97
  ```bibtex
98
+ @article{russell2026wildai,
99
  title = {How Much Is an AI Token Worth? Scaling Laws for Wild AI-Generated Web Text},
100
  author = {Russell, Jenna and Glickenhaus, Ben and Thai, Katherine and Wieting, John and Iyyer, Mohit and Spero, Max and Emi, Bradley},
101
  journal = {arXiv preprint arXiv:2609.40295},
35.8m-g024-ai-r1/README.md CHANGED
@@ -18,7 +18,7 @@ tags:
18
 
19
  > **License: [CC BY-NC-SA 4.0](https://creativecommons.org/licenses/by-nc-sa/4.0/).** The WildAI code, models and dataset are released under the Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International license: you may use and adapt them for non-commercial purposes, with attribution, and must share adaptations under the same license.
20
 
21
- A 35.8M decoder-only language model trained from scratch for [*How Much Is an AI Token Worth? Scaling Laws for Wild AI-Generated Web Text*](https://arxiv.org/abs/2609.40295), one of 844 models that measure how wild AI-generated web text in pretraining data changes a model's loss. It is a research artifact for studying scaling laws, not an assistant: it has had no instruction tuning or safety training.
22
 
23
  ## Training data
24
 
@@ -95,7 +95,7 @@ The nanochat GPT recipe: rotary embeddings, RMSNorm without weights (also on que
95
  ## Citation
96
 
97
  ```bibtex
98
- @article{russell2026aitoken,
99
  title = {How Much Is an AI Token Worth? Scaling Laws for Wild AI-Generated Web Text},
100
  author = {Russell, Jenna and Glickenhaus, Ben and Thai, Katherine and Wieting, John and Iyyer, Mohit and Spero, Max and Emi, Bradley},
101
  journal = {arXiv preprint arXiv:2609.40295},
 
18
 
19
  > **License: [CC BY-NC-SA 4.0](https://creativecommons.org/licenses/by-nc-sa/4.0/).** The WildAI code, models and dataset are released under the Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International license: you may use and adapt them for non-commercial purposes, with attribution, and must share adaptations under the same license.
20
 
21
+ A 35.8M decoder-only language model trained from scratch for [*How Much Is an AI Token Worth? Scaling Laws for Wild AI-Generated Web Text*](https://arxiv.org/abs/2609.40295), one of the paper's models measuring how wild AI-generated web text in pretraining data changes a model's loss. It is a research artifact for studying scaling laws, not an assistant: it has had no instruction tuning or safety training.
22
 
23
  ## Training data
24
 
 
95
  ## Citation
96
 
97
  ```bibtex
98
+ @article{russell2026wildai,
99
  title = {How Much Is an AI Token Worth? Scaling Laws for Wild AI-Generated Web Text},
100
  author = {Russell, Jenna and Glickenhaus, Ben and Thai, Katherine and Wieting, John and Iyyer, Mohit and Spero, Max and Emi, Bradley},
101
  journal = {arXiv preprint arXiv:2609.40295},
35.8m-g024-control/README.md CHANGED
@@ -18,7 +18,7 @@ tags:
18
 
19
  > **License: [CC BY-NC-SA 4.0](https://creativecommons.org/licenses/by-nc-sa/4.0/).** The WildAI code, models and dataset are released under the Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International license: you may use and adapt them for non-commercial purposes, with attribution, and must share adaptations under the same license.
20
 
21
- A 35.8M decoder-only language model trained from scratch for [*How Much Is an AI Token Worth? Scaling Laws for Wild AI-Generated Web Text*](https://arxiv.org/abs/2609.40295), one of 844 models that measure how wild AI-generated web text in pretraining data changes a model's loss. It is a research artifact for studying scaling laws, not an assistant: it has had no instruction tuning or safety training.
22
 
23
  ## Training data
24
 
@@ -95,7 +95,7 @@ The nanochat GPT recipe: rotary embeddings, RMSNorm without weights (also on que
95
  ## Citation
96
 
97
  ```bibtex
98
- @article{russell2026aitoken,
99
  title = {How Much Is an AI Token Worth? Scaling Laws for Wild AI-Generated Web Text},
100
  author = {Russell, Jenna and Glickenhaus, Ben and Thai, Katherine and Wieting, John and Iyyer, Mohit and Spero, Max and Emi, Bradley},
101
  journal = {arXiv preprint arXiv:2609.40295},
 
18
 
19
  > **License: [CC BY-NC-SA 4.0](https://creativecommons.org/licenses/by-nc-sa/4.0/).** The WildAI code, models and dataset are released under the Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International license: you may use and adapt them for non-commercial purposes, with attribution, and must share adaptations under the same license.
20
 
21
+ A 35.8M decoder-only language model trained from scratch for [*How Much Is an AI Token Worth? Scaling Laws for Wild AI-Generated Web Text*](https://arxiv.org/abs/2609.40295), one of the paper's models measuring how wild AI-generated web text in pretraining data changes a model's loss. It is a research artifact for studying scaling laws, not an assistant: it has had no instruction tuning or safety training.
22
 
23
  ## Training data
24
 
 
95
  ## Citation
96
 
97
  ```bibtex
98
+ @article{russell2026wildai,
99
  title = {How Much Is an AI Token Worth? Scaling Laws for Wild AI-Generated Web Text},
100
  author = {Russell, Jenna and Glickenhaus, Ben and Thai, Katherine and Wieting, John and Iyyer, Mohit and Spero, Max and Emi, Bradley},
101
  journal = {arXiv preprint arXiv:2609.40295},
35.8m-g025-ai-r0.0256/README.md CHANGED
@@ -18,7 +18,7 @@ tags:
18
 
19
  > **License: [CC BY-NC-SA 4.0](https://creativecommons.org/licenses/by-nc-sa/4.0/).** The WildAI code, models and dataset are released under the Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International license: you may use and adapt them for non-commercial purposes, with attribution, and must share adaptations under the same license.
20
 
21
- A 35.8M decoder-only language model trained from scratch for [*How Much Is an AI Token Worth? Scaling Laws for Wild AI-Generated Web Text*](https://arxiv.org/abs/2609.40295), one of 844 models that measure how wild AI-generated web text in pretraining data changes a model's loss. It is a research artifact for studying scaling laws, not an assistant: it has had no instruction tuning or safety training.
22
 
23
  ## Training data
24
 
@@ -95,7 +95,7 @@ The nanochat GPT recipe: rotary embeddings, RMSNorm without weights (also on que
95
  ## Citation
96
 
97
  ```bibtex
98
- @article{russell2026aitoken,
99
  title = {How Much Is an AI Token Worth? Scaling Laws for Wild AI-Generated Web Text},
100
  author = {Russell, Jenna and Glickenhaus, Ben and Thai, Katherine and Wieting, John and Iyyer, Mohit and Spero, Max and Emi, Bradley},
101
  journal = {arXiv preprint arXiv:2609.40295},
 
18
 
19
  > **License: [CC BY-NC-SA 4.0](https://creativecommons.org/licenses/by-nc-sa/4.0/).** The WildAI code, models and dataset are released under the Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International license: you may use and adapt them for non-commercial purposes, with attribution, and must share adaptations under the same license.
20
 
21
+ A 35.8M decoder-only language model trained from scratch for [*How Much Is an AI Token Worth? Scaling Laws for Wild AI-Generated Web Text*](https://arxiv.org/abs/2609.40295), one of the paper's models measuring how wild AI-generated web text in pretraining data changes a model's loss. It is a research artifact for studying scaling laws, not an assistant: it has had no instruction tuning or safety training.
22
 
23
  ## Training data
24
 
 
95
  ## Citation
96
 
97
  ```bibtex
98
+ @article{russell2026wildai,
99
  title = {How Much Is an AI Token Worth? Scaling Laws for Wild AI-Generated Web Text},
100
  author = {Russell, Jenna and Glickenhaus, Ben and Thai, Katherine and Wieting, John and Iyyer, Mohit and Spero, Max and Emi, Bradley},
101
  journal = {arXiv preprint arXiv:2609.40295},
35.8m-g025-ai-r0.0514/README.md CHANGED
@@ -18,7 +18,7 @@ tags:
18
 
19
  > **License: [CC BY-NC-SA 4.0](https://creativecommons.org/licenses/by-nc-sa/4.0/).** The WildAI code, models and dataset are released under the Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International license: you may use and adapt them for non-commercial purposes, with attribution, and must share adaptations under the same license.
20
 
21
- A 35.8M decoder-only language model trained from scratch for [*How Much Is an AI Token Worth? Scaling Laws for Wild AI-Generated Web Text*](https://arxiv.org/abs/2609.40295), one of 844 models that measure how wild AI-generated web text in pretraining data changes a model's loss. It is a research artifact for studying scaling laws, not an assistant: it has had no instruction tuning or safety training.
22
 
23
  ## Training data
24
 
@@ -95,7 +95,7 @@ The nanochat GPT recipe: rotary embeddings, RMSNorm without weights (also on que
95
  ## Citation
96
 
97
  ```bibtex
98
- @article{russell2026aitoken,
99
  title = {How Much Is an AI Token Worth? Scaling Laws for Wild AI-Generated Web Text},
100
  author = {Russell, Jenna and Glickenhaus, Ben and Thai, Katherine and Wieting, John and Iyyer, Mohit and Spero, Max and Emi, Bradley},
101
  journal = {arXiv preprint arXiv:2609.40295},
 
18
 
19
  > **License: [CC BY-NC-SA 4.0](https://creativecommons.org/licenses/by-nc-sa/4.0/).** The WildAI code, models and dataset are released under the Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International license: you may use and adapt them for non-commercial purposes, with attribution, and must share adaptations under the same license.
20
 
21
+ A 35.8M decoder-only language model trained from scratch for [*How Much Is an AI Token Worth? Scaling Laws for Wild AI-Generated Web Text*](https://arxiv.org/abs/2609.40295), one of the paper's models measuring how wild AI-generated web text in pretraining data changes a model's loss. It is a research artifact for studying scaling laws, not an assistant: it has had no instruction tuning or safety training.
22
 
23
  ## Training data
24
 
 
95
  ## Citation
96
 
97
  ```bibtex
98
+ @article{russell2026wildai,
99
  title = {How Much Is an AI Token Worth? Scaling Laws for Wild AI-Generated Web Text},
100
  author = {Russell, Jenna and Glickenhaus, Ben and Thai, Katherine and Wieting, John and Iyyer, Mohit and Spero, Max and Emi, Bradley},
101
  journal = {arXiv preprint arXiv:2609.40295},