phase3
Datasets
All datasets matching “phase3”phase3-redraft
Phase 3: re-draft between judge passes
SDAR-1.7B drafter + Phase 1's 128M LoRA judge on JetLM/SDAR-30B-A3B-Chat-b32 (block 32, threshold 0.95, no expert pool), re-drafting the [MASK] slots from committed tokens after every judge pass (rd1) or every 2 passes (rd2), on 7 benchmarks with one 8x RTX PRO 6000 box.
code: github.com/NoviceCoderInfinity/diffusion-moe-general-expert, branch phase3-redraft, commit 2ebaba7762f4ae44bb8a49349ad570f762f64326
results/phase3/TABLE.md: the… See the full description on the dataset page: https://huggingface.co/datasets/Arushhh/phase3-redraft.phase3-layer0-persona-direction-qwen3-8b
phase3-layer0-persona-direction — Qwen3-8B, 10,560 rollouts under a layer-0 persona direction
Generations and residual-stream activations from the "brrrt" run of phase 3 of
mech-interp-on-randomly-emergent-personas
(the follow-on to CoT-spiking). Phase 3 fits a single direction
added to the residual stream after block 0 of Qwen/Qwen3-8B so that the layer-20 first-answer state moves
off phase 1's INLP-debiased assistant axis. On unseen prompts the model answers in a persona… See the full description on the dataset page: https://huggingface.co/datasets/mild-rgb/phase3-layer0-persona-direction-qwen3-8b.phase3-dataset
Dataset
Generated dataset with 120 configurations.
Configuration
dataset_name: phase3-dataset
graphs_path: hf://CSE472-blanket-challenge/phase3-graphs
output_path: data/datasets/${dataset_name}
n_samples: 1000
n_datasets: 1
scm_type: linear, nonlinear
coeff_range: 1.0
noise_std: 0.5
env_type: iid, covariate, label
projection: pca
shift_mean: 0.8
shift_std: 0.2
train_fraction: 0.8
seed: 42
overwrite: true
Load data
from huggingface_hub import snapshot_download… See the full description on the dataset page: https://huggingface.co/datasets/CSE472-blanket-challenge/phase3-dataset.LinguaFranca-Phase3origami-phase3-mixed-speed-shardslensemble-phase3-so100-silos
Lensemble Phase 3 — SO-100 consortium silos + held-out eval split
Deterministic episode-modulo split (k % 5) of the public SO-100 pick-place set
abdelstark/so100-pickplace-lewm-ready
into four sovereign participant silos plus one disjoint held-out eval split, for the Phase 3
federated JEPA / LeWorldModel consortium run (Lensemble epic #249, issue #242).
file
role
episodes
windows (window_steps=4)
dataset Merkle root (sha256, prefix)
phase3-so100-silo0.h5
participant… See the full description on the dataset page: https://huggingface.co/datasets/abdelstark/lensemble-phase3-so100-silos.