Few-GT step-300 output directions
This repository holds compact, per-matrix rank-one left/output singular
vectors of W_GT300 - W0, from true-reward RL with 64, 256 or 512 distinct
DeepScaleR labels. Files use {"metadata": ..., "bases": {parameter: u1}}.
They are direction tensors, not model weights, adapters or input/right bases.
manifest.json is the authoritative inventory. Qwen N64/N256/N512 and Llama
N64/N512 GT300 artifacts are supplied initially. Only Llama N256 remains
missing until its GT300 donor is trained and extracted. Never infer
availability or step counts from this card alone.
Qwen targets196 matrices; Llama targets112. Every matrix is one of
q/k/v/o/gate/up/down projections. Vectors are FP32 unit columns. The original
Qwen d0 SVD caches stored FP16 factors; their first U column is converted to
FP32 and normalized, without recomputing the SVD. Artifact metadata records
source identities, cache SHA256 hashes, exact base revisions and limitations.
Full historical donor weights are not part of this repository.
Llama N64/N512 directions were extracted from archived fewgold_llama_N*_c300
checkpoints using FP32 differences from the pinned base and
torch.svd_lowrank(q=4,niter=4). Their metadata records checkpoint and training
log hashes. Llama N64 uses the earlier fewgold prompt selection, which differs
from the N64 d0 pool; Llama N512 uses fewgold_N512_d0. Label counts match
the stated budgets, but prompt selections need not match across model families.
The associated portable training package initializes and anchors guided runs
at W₀, with lambda/2 * sum ||(I-u1*u1^T)(W-W0)||_F^2 and label-free
majority-vote GRPO. These files do not supply a W_GT initialization or anchor.
Respect the applicable upstream model terms when using or redistributing model-derived artifacts. Source code in the accompanying verl package retains its own license. No base-model weights, raw training data or credentials are uploaded by the supplied upload helper.