fed-pulse multi-axis text classifier
Per-seed text classifier checkpoints from the fed-pulse §6.41 wave-4
encoder bake-off (see 06_Deep_Learning_Roadmap.md in the wiki at
https://github.com/yusufizzetmurat/fed-pulse.wiki).
Each text_multi_axis_seedNN.pt is a single-seed fine-tuned
finbert_fed_adjacent encoder plus a multi-task classification head
(stance / factor / certainty). text_multi_axis_best.pt is the
singleton inference checkpoint the /analyze service loads from
backend/models/text_multi_axis_best.pt.
Seeds 11, 47, 97 are checked in here; seeds 29 and 71 were trained pod-side only and are not in this repo. The §6.41 5-row encoder bake-off table reports the canonical 5-seed mean computed at training time — the 3 local checkpoints suffice to reproduce inference; the missing 2 seeds were used for the §6 numbers but were not retained as artefacts post-training.
- Source code: https://github.com/yusufizzetmurat/fed-pulse
- Encoder backbone:
yusufizzetmurat/finbert-fed-adjacent(rev 332e31e166) - License: MIT