Download README.md from AwakeningOS/vistarc-research-notes: direct link, hf CLI and curl.
- Browser
- Download file 2.37 kB
-
https://huggingface.co/spaces/AwakeningOS/vistarc-research-notes/resolve/main/README.md
- Command line
-
hf download hf://spaces/AwakeningOS/vistarc-research-notes/README.md
-
curl -L -o README.md https://huggingface.co/spaces/AwakeningOS/vistarc-research-notes/resolve/main/README.md
title: VISTARC Research Notes
emoji: 🧭
colorFrom: green
colorTo: blue
sdk: static
app_file: index.html
pinned: false
license: apache-2.0
VISTARC Research Notes / VISTARC研究ノート
A bilingual research notebook from AwakeningOS: 13 explanatory chapters, a searchable and sortable catalogue of 302 permanent experimental identities, 988 evaluation points, and a frozen 17,274-item clear-answer benchmark evaluated on 16 models.
Start with What this research investigates and How to read scores and comparisons in the app. No knowledge of the original experiment conversations is required. Model IDs are permanent identities, not ranks.
The notebook explains architecture components, seeds, learning exposure, separate observed five-/six-axis peaks, pooled clear-answer accuracy, and the limits of causal conclusions from the available experiments. It includes unsuccessful and missing results. Legacy rankings are separated by comparison cohort.
Data and provenance
data/models.csv summarizes each identity. data/checkpoints.csv includes every recorded evaluation point. JSON files expose configurations, effective named learning rates/weight decay, scoring protocol, and source identifiers/hashes. data/manifest.json records the snapshot and coverage. SHA256SUMS verifies this public package. Source references do not imply that all original source documents are redistributed. Question text, training data, weights, secrets, and private working paths are not included.
The clear-answer bank is a research diagnostic screened with an API. It is not an official contest ranking, and screening does not establish that every question is unambiguous. Seven-task supervised fine-tuning results are not mixed into the five-/six-axis or clear-answer rankings here.
Released model
VISTARC-320-9L-MoS3-80M, including English and Japanese architecture guides, weights, code, and evaluation documentation.
Reuse
Notebook text, presentation code, and original aggregate records: Apache-2.0. Referenced external datasets and projects retain their own terms. Cite AwakeningOS, VISTARC Research Notes, snapshot 2026-11-02, and the repository revision you used. This is an observational research log rather than a peer-reviewed claim of universal architecture superiority.