| Barbados — Third-party attribution and distribution notices |
|
|
| Barbados is an independent research prototype published by Vykos. The uploader |
| licenses the original model weights and project-authored runtime under the |
| Apache License, Version 2.0, reproduced in LICENSE. Bundled third-party works |
| retain their licenses, copyrights and notices. No affiliation or endorsement |
| by Mistral AI, the cited paper authors or dataset contributors is implied. |
|
|
| LOOM source |
| |
| Source: hed-ucas/LOOM |
| https://github.com/hed-ucas/LOOM/tree/a2ee5a7ecd38b083d9c049d97271ece38f4bce4c |
| Pinned commit: a2ee5a7ecd38b083d9c049d97271ece38f4bce4c |
| License: Apache-2.0; the upstream text is retained at bundle/source/LICENSE. |
| Original upstream attribution, comments and file contents are preserved except |
| for the project modifications recorded in the trained artifact and the |
| nonexecuting public-distribution notices described below. |
|
|
| These six files contain project modifications relative to the pinned upstream: |
| bundle/source/models/baselines/looped_transformer.py |
| bundle/source/models/layers.py |
| bundle/source/models/lm_head.py |
| bundle/source/models/moe.py |
| bundle/source/models/transformer.py |
| bundle/source/pretrain.py |
| The distribution adds prominent change comments to those files. The original |
| project extension bundle/source/models/melt_memory.py receives an Apache-2.0 |
| license comment. All seven edits are comments only relative to the completed |
| trained artifact. ASTs of all 34 Python source files are unchanged, and all |
| 19 non-Python source files remain byte-identical. Runtime integrity constants |
| and the distribution receipts are rebound without removing any hash guard. |
| Model weights, tokenizer bytes and inference algorithms are unchanged. |
| Original-to-public hashes and exact original receipts are retained in release/. |
|
|
| Tokenizer |
| |
| HuggingFaceTB/SmolLM-360M tokenizer, Apache-2.0 according to its pinned model card: |
| https://huggingface.co/HuggingFaceTB/SmolLM-360M/tree/59f7ef243ee09a72cbc14cb054393a3e3b771d41 |
| Revision: 59f7ef243ee09a72cbc14cb054393a3e3b771d41 |
| Tokenizer JSON SHA-256: |
| 9ca9acddb6525a194ec8ac7a87f24fbba7232a9a15ffa1af0c1224fcd888e47c |
| The six reused tokenizer/configuration files are bundled unchanged. The SmolLM |
| foundation weights were not used to initialize Barbados and are not included. |
| Apache-2.0 is reproduced in LICENSE; existing tokenizer metadata is retained. |
|
|
| Pretraining data attribution — data not redistributed |
| |
| HuggingFaceFW and the FineWeb / FineWeb-Edu dataset creators: |
| https://huggingface.co/datasets/HuggingFaceFW/fineweb-edu/tree/87f09149ef4734204d70ed1d046ddc9ca3f2b8f9 |
| Revision: 87f09149ef4734204d70ed1d046ddc9ca3f2b8f9 |
| Selected sample/10BT files: 013_00000.parquet and 000_00000.parquet. |
| The dataset card declares Open Data Commons Attribution License (ODC-By) v1.0: |
| https://opendatacommons.org/licenses/by/1-0/ |
| Use is also subject to CommonCrawl's Terms of Use: |
| https://commoncrawl.org/terms-of-use |
| These dataset terms are separate from the model-weight/code license. This |
| release does not redistribute the raw documents, tokenized training corpus or |
| training examples. Provenance receipts and scientific evaluation statistics |
| are included. Dataset curation does not establish rights in every source text. |
| |
| Visible-response SFT data attribution — data not redistributed |
| ------------------------------------------------------------- |
| TheFusionCube/Fable-5-CoT-Traces: |
| https://huggingface.co/datasets/TheFusionCube/Fable-5-CoT-Traces/tree/3624e3a8aab36a89dc3b89aebef0e034a64ea04e |
| Revision: 3624e3a8aab36a89dc3b89aebef0e034a64ea04e |
| The uploader's pinned dataset card declares MIT. The selected data comprise |
| public visible explanatory responses; no authenticated hidden/private |
| chain-of-thought is claimed. The generating model's identity and response |
| correctness are not independently authenticated. The uploader's license |
| label does not establish rights in every response. The duplicate |
| armand0e/Fable-5-Chat conversion contributes no additional training examples. |
| No source responses or SFT examples are redistributed in this release. |
|
|
| Research acknowledgements |
| |
| LOOM: https://arxiv.org/abs/2610.01153v1 |
| Ouro: https://arxiv.org/abs/2510.25741v5 |
| MELT: https://arxiv.org/abs/2605.07721v3 |
| This reduced single-seed experiment is not a full-scale reproduction of any |
| of these published training runs. Existing papers supply the recurring-stack, |
| sparse-expert and latent-memory-sharing ideas acknowledged in README.md. |
|
|
| PIQA evaluation annotations and derived measurements |
| |
| Source: PIQA dataset project, ybisk/ybisk.github.io/piqa/data |
| https://github.com/ybisk/ybisk.github.io/tree/16c6be5562fdf2e56a6d5ee4f39a047494eb0617/piqa/data |
| Data revision: 16c6be5562fdf2e56a6d5ee4f39a047494eb0617 |
| The current upstream README declares Academic Free License v3.0 (AFL-3.0): |
| https://github.com/ybisk/ybisk.github.io/blob/master/piqa/README.md |
| That README is reproduced unchanged in licenses/PIQA-upstream-README.md; |
| the standard license is reproduced in licenses/PIQA-AFL-3.0.txt. The README |
| is a current license declaration, not a file present at the old data revision. |
| PIQA annotations retain their applicable upstream terms, separate from the |
| Apache-2.0 license for original model weights and project-authored code. |
|
|
| Modification/attribution notice: evaluation outputs retain original case IDs and |
| answer labels and add project-computed predictions, losses, aggregates and |
| numeric routing diagnostics. Raw questions, candidate strings and vocabulary |
| token IDs are not redistributed. No PIQA data is used for new training or tuning |
| in these evaluation runs. The public validation set has now been fully scored, |
| including the 1,338 rows previously reserved locally for final evaluation. |
|
|
