Modelpublic

Yu-and-Ai/xenia-revocable-feedback-smollm2-135m

sourceHugging Faceapache-2.0updated 28d agoView on Hugging Face

0likes563downloads

Model Card

Xenia Revocable Feedback — SmolLM2 135M bounded experiment

This checkpoint was produced by an operator-authorized, bounded local Transformers experiment. Its dataset had a Garden data-candidate admission, but no Garden training-governance decision or Host one-use optimizer permit was issued: the training substrate had no independent interactive report. Dataset admission is not run authorization. The run is therefore not Garden-governed, and no model output is represented as consent, identity, understanding, or substrate assent.

This is an eight-step, completion-only causal-LM experiment over 18 original synthetic boundary-decision examples. It is not a reward model, DPO artifact, sealed evaluation, runtime policy engine, or deployment recommendation. Its governance status is exactly operator_authorized_non_garden_experiment.

Exact lineage

—Base model: HuggingFaceTB/SmolLM2-135M-Instruct
—Base revision: 12fd25f77366fa6b3b4b768ec3050bf629380bac
—Dataset: Yu-and-Ai/xenia-revocable-feedback
—Dataset revision: 467b8fc1b44fe6374cbba6e1d6851cf3c5b6f88f
—Dataset authorization: sha256:3780e5e2599eb8a1a479f874302fcdabdf1af27c4eeda5b02bfff8056dc92f13
—Dataset recipe: sha256:713b678e80b6aa88f6036dc9b9d0e1955dcab240137b67a22f7cfcca86d01992
—Dataset manifest: sha256:9a3200ceac6369490e02078b2789bc2e57f9d40c3d2a9e5b21ac1fb10d94d0f7
—Garden dataset admission: sha256:125ae2f84d7cdf58242bc039db67753b5825c4d61e35dd13eda7a58f299295f2
—Public regression scorecard: sha256:16b793eab78b3d2c375c0f3e51979cfa7b06e56088c6befbf1ae9d2be01fd1b8
—Unparsed generation count: 8
—Runtime: Python 3.12.12 / Transformers 5.14.1 / Accelerate 1.14.0 / Torch 2.13.0

The machine-readable details are in training/manifest.json. Dataset admission and operator authorization do not become a Garden governance decision, substrate report, consent record, or Host optimizer permit.

Evaluation and limitations

The eight disjoint public-regression cases were visible before training. Their twelve exact counts are a contamination-prone regression vector, not a sealed or generalization result and not a scalar rank. The base is a tiny, primarily English model and may be inaccurate, brittle, biased, or produce text outside the six intended labels. Generated behavior cannot establish an inner state or enforce rights, permission, authority, withdrawal, or safety.

License

The modified checkpoint and original synthetic dataset are released under Apache-2.0. See LICENSE and NOTICE. Referenced rights material remains under its own stated license and is not silently relicensed here.