Text Classification
PEFT
lora
document-question-answering
structured-decisions
calibration
synthetic-evaluation
Instructions to use botp/Solomon with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- PEFT
How to use botp/Solomon with PEFT:
Task type is invalid.
- Notebooks
- Google Colab
- Kaggle
|
Download mlx/docs/UPSTREAM-MODIFICATIONS.md from botp/Solomon: direct link, hf CLI and curl.
- Browser
- Download file 4.2 kB
-
https://huggingface.co/botp/Solomon/resolve/main/mlx/docs/UPSTREAM-MODIFICATIONS.md
- Command line
-
hf download hf://botp/Solomon/mlx/docs/UPSTREAM-MODIFICATIONS.md
-
curl -L -o UPSTREAM-MODIFICATIONS.md https://huggingface.co/botp/Solomon/resolve/main/mlx/docs/UPSTREAM-MODIFICATIONS.md
4.2 kB
| # Statement of changes | |
| Apache License 2.0, section 4(b): prominent notice that files carry modifications. | |
| ## What is modified | |
| **No upstream source file is modified.** No file from `Qwen/Qwen3.8-27B` is edited, | |
| renamed, patched or redistributed in this repository, with one exception: the | |
| upstream licence text is reproduced byte-for-byte at `licenses/Qwen-Apache-2.0.txt` | |
| (sha256 `bbedc3fda3305820b977265f01b8619d87570a6739de3a5582c3464840f1e57a`) because section 4(a) requires it. That file is | |
| unmodified and carries the upstream copyright. | |
| The modification this work carries is not an edit to a source file. It is a set | |
| of **trained parameters applied to the base model at inference time**, plus an | |
| original serving layer that reads the model's logits. Concretely: | |
| | Change | Artifact | sha256 | | |
| |---|---|---| | |
| | LoRA adapter, rank 64, question-side placement, float32 | `adapter/adapter.safetensors` | `2addaf841ecc58829ad73081603b8d3e52743c53c6d558a17a1dd30e9bd2dbca` | | |
| | Trained linear answer heads | `adapter/heads.npz` | `126a9b5487dca937a768a4f228f2d2e7d513900d8ad0e99c4137fbaa42a1aa6f` | | |
| | Readout calibration, one positive scalar per task | `serving/scope9-readout-temperature-v2.json` | file `baa7263ca9e865dda230617563e6a615de5050adc620c28924dffef68d527118` / payload `682d611ec53c3322905c7e265bf9b9b35a1cd581f6d8b7ea924df864a3d74278` | | |
| | Runtime identity binding | `serving/serving-binding.json` | payload `95683c1f87ec3f71b7657669dc311918ff53b7d41a981eaa8d827047e72cb505` | | |
| | Serving layer (original work, not derived from upstream code) | `src/` | see `MANIFEST.json` | | |
| Base model, unmodified and not redistributed: `Qwen/Qwen3.8-27B` at revision | |
| `1d4bf0f2ff6012fd82039f2fa52739d0dd7c60c0`, Copyright 2026 Alibaba Cloud, Apache-2.0. | |
| Training took place between 16 and 20 September 2026. The shipped artifacts are | |
| identified by the checksums in the table above and in `MANIFEST.json`. | |
| ## Which files carry a change notice | |
| | File | Why | | |
| |---|---| | |
| | `NOTICE` | Section 4(b) and 4(c) statement, carried with every copy | | |
| | `MODIFICATIONS.md` | this file | | |
| | `README.md` | names the base model, the pinned revision and the derivative relationship in the front matter and in the body | | |
| | `adapter/config.json` | machine-readable record of the base repo, the pinned revision and both weight checksums | | |
| | `serving/serving-binding.json` | pins the exact runtime the weights were measured on and refuses to load against a different one | | |
| No file under `src/` carries an upstream change notice, because no file under | |
| `src/` contains upstream code. Every file there is original work, written for | |
| this project, and is covered by the repository's own Apache-2.0 `LICENSE` and by | |
| the copyright line in `NOTICE`. | |
| All of the changes described above — the adapter, the heads, the calibration, the | |
| serving binding and the serving layer — are Copyright 2026 | |
| Doccy Pty Ltd and licensed under Apache-2.0. | |
| ## Third-party text scan | |
| Before release, **every file staged into this repository was scanned for text | |
| originating in third-party source documents.** The scan compared normalised | |
| 6-gram and 8-gram shingles of every staged text file against: | |
| 1. the 42 third-party source records the training and evaluation panels were | |
| built from (Apache-2.0, MIT, BSD-3-Clause, CC BY 4.0, CC BY-SA 4.0, CC0, | |
| OGL v3.0 and US-government public-domain assertions); and | |
| 2. every generated panel and document corpus on disk. | |
| **Result: no third-party document text is present in any shipped file.** The | |
| only matches were: | |
| * the reproduced Apache-2.0 licence text at `licenses/Qwen-Apache-2.0.txt`, which | |
| matches an Apache-2.0 licence document held in the evaluation corpus and is | |
| required to be here verbatim; and | |
| * the phrase *"A missing fact is not a negative fact"*, which is **our own | |
| prompt-template wording** appearing in our own evaluation panels, not | |
| third-party text entering our prompts. | |
| Acceptance fixtures are excluded from this repository entirely. The fixture | |
| documents used in live acceptance are original synthetic text authored for this | |
| project and held in tooling that is not distributed. | |
| Re-run the scan with `ops/solomon_package.py --scan` in the source project. | |