| # Attribution | |
| Our Oracle MVP subject-extraction adapter contributions and original project | |
| material, including the ASv6.100 prompt, are licensed under GPL-3.0-or-later | |
| (LICENSE). Copyright 2026 kernelkind. | |
| The base models retain Apache-2.0 licensing. The generic license text is in | |
| LICENSES/Apache-2.0.txt; Qwen’s original license and copyright notice are preserved | |
| in LICENSES/Qwen-LICENSE. Base models and tokenizers remain upstream dependencies. | |
| - Qwen/Qwen3.8-27B, revision `1d4bf0f2ff6012fd82039f2fa52739d0dd7c60c0`: | |
| https://huggingface.co/Qwen/Qwen3.8-27B/tree/1d4bf0f2ff6012fd82039f2fa52739d0dd7c60c0 | |
| - google/gemma-4-31B-it, revision `842da3794eaa0b77d5f08bae87a17459d91ff475`: | |
| https://huggingface.co/google/gemma-4-31B-it/tree/842da3794eaa0b77d5f08bae87a17459d91ff475 | |
| - google/gemma-4-26B-A4B-it, revision `4d7ae4984b7db7de8f8457170b3f1a419ee76d52`: | |
| https://huggingface.co/google/gemma-4-26B-A4B-it/tree/4d7ae4984b7db7de8f8457170b3f1a419ee76d52 | |
| The saved research adapter files are distributed unchanged. Their training | |
| history and hashes are in each family’s checkpoints.json. These are trained | |
| LoRA additions rather than merged upstream weights. No upstream endorsement | |
| is implied. | |