966Aurora/continuous-av-scorers

Frozen continuous A/V research scorers

This repository bundles frozen f1 and f2 inference artifacts for a continuous affective-expression selection study. It is a multi-scorer archive, not a single model loadable from its root with from_pretrained.

Contents

Path	Meaning
f1/	Three frozen RoBERTa V/A/D regression members, tokenizers, manifest and selection results
f2/encoder/	Pinned sentence-transformers/all-mpnet-base-v2 encoder
f2/head/	Three PAD ridge heads, feature centers and baselines
f3/upstream_reference.json	Reference to the external Mendes & Martins XLM-RoBERTa-base MSE checkpoint; no f3 weights included
validation.json	Re-inference validation and scorer identities
files_sha256.json	Sizes and SHA256 hashes of uploaded files

No receiver database, API keys, requests, training corpus, optimizer state or unselected training checkpoints is included.

f1

Reader-perspective V/A/D regression trained using EmoBank. Three members are averaged equally. Seeds 20261003, 20261004 and 20261005 use selected epochs 1, 2 and 2, respectively; selection was by development A RMSE. Dimensions are V, A, D and scores use the original 1–5 scale. The supplied result records test_evaluated; development validation must not be described as an independent test result.

Frozen scorer revision: 9fc2f9db01ad0f608e01bdab4fa1c41ef427514d003a30fe2d07508bfe5512a1.

f2

Frozen MPNet commit e8c3b32edf5434bc2275fc9bab85f82640a19130. Encoding uses attention-mask mean pooling including special tokens, L2 normalization and FP32. Each PAD head predicts (embedding - center) @ weights + baseline; baselines are in f2/head/result.json. The ridge heads are project artifacts; the encoder is an unchanged upstream model. Both f1 and f2 use EmoBank supervision and are not independent human ground truths.

Restore into the handoff package

Download this repository at a fixed commit, then copy:

f1/*          -> continuous_AV_handoff_20261015/workspace/results/f1_official_tuned_v1/
f2/encoder/*  -> continuous_AV_handoff_20261015/workspace/models/all_mpnet_base_v2/
f2/head/*     -> continuous_AV_handoff_20261015/workspace/results/absolute_pad_v2/f2/

Use the handoff package's rescore.py and requirements-scoring.txt. SHA256 identities must match before inference. The historical absolute_pad_v2 path is retained for compatibility; it does not merge a different study's outcomes into the continuous AV study.

All 534 frozen scoring inputs were re-scored on the validation machine on 2026-10-15; f1 and f2 matched the archived values exactly. Cross-device or precision changes may yield small numerical differences. This is inference reproducibility evidence, not a new construct-validity evaluation.

f3

The external model is Mendes & Martins multilingual VA prediction, XLM-RoBERTa-base MSE. It outputs V then A, on 0–1, with the author's custom regression head and hard-sigmoid output. It has no D dimension and should not be loaded as an ordinary softmax classifier. This local project received scores and asset hashes from remote execution, not the checkpoint files. Exact source URLs and hashes are preserved in f3/upstream_reference.json.

The author code repository has an MIT notice. This archive does not assume that the notice unambiguously licenses separately hosted checkpoint files for mirroring; use the author's download link unless checkpoint redistribution is confirmed. f3 is a post-experiment sensitivity scorer, not the original primary measurement.

Scope and limitations

These are computational text affect estimates, not evidence of a language model's subjective emotions. Scores can depend on wording, context and domain. In the associated study the ten-event focus is a post-outcome subset of the original twelve events; this model release does not turn that analysis into a preregistered result. No universal independent AV causal effect is claimed.

Sources and notices

• RoBERTa: https://huggingface.co/FacebookAI/roberta-base
• MPNet model card and Apache-2.0 notice: https://huggingface.co/sentence-transformers/all-mpnet-base-v2
• EmoBank reader labels, CC-BY-SA-4.0 dataset notice and citations: https://github.com/JULIELab/EmoBank
• Buechel & Hahn (2017), EmoBank: Studying the Impact of Annotation Perspective and Representation Format on Dimensional Emotion Analysis: https://aclanthology.org/E17-2092/
• Mendes & Martins source code: https://github.com/gmendes9/multilingual_va_prediction

Component and source notices are recorded separately. No blanket new license for project-trained weights is asserted by this archive. Dataset notices are not automatically represented as a legal determination of checkpoint licensing. The initial intended use is private research handoff.