YAML Metadata Warning:empty or missing yaml metadata in repo card
Check out the documentation for more information.
DeepSeek-V4-Flash — early-lens & J-lens artifacts
Jacobian lenses for DeepSeek-V4-Flash (~280B FP8/FP4 MoE, mHC residual), fit on 25 pile
prompts, target_layer=-2 (penultimate), skip_first=4, max_seq_len=128; merged
per-prompt partials from Modal volume jlens-ablations
(outputs/deepseek-v4-flash/<arm>/n25/lens.pt). Layout: <model>/<arm>-n25/lens.pt.
| file | arm | gradient rules |
|---|---|---|
deepseek-v4-flash/j-lens-n25/lens.pt |
std |
standard gradients — the J-lens baseline |
deepseek-v4-flash/early-lens-all-c4-n25/lens.pt |
all-c4 |
early-lens: LN-rule + identity-rule + half-rule, routed+shared experts as MLPs, mHC coefficients detached, router frozen, shared-expert grad-scale c=4 — sweep winner |
See docs/project/experiments/relp_jlens/replication.md (§DeepSeek-V4-Flash) in
global-workspace for the detach set and eval protocol.
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support