Cross-transliteration Tibetan embedding work.
Anna Maresova PRO
anicka
AI & ML interests
Cybersecurity, AI alignment
Recent Activity
updated a collection 5 days ago
Tibetan NLP updated a collection 5 days ago
CVE backport codegen updated a collection 5 days ago
CVE backport codegenOrganizations
Steering vectors and probes from the valence-readout / introspection-frame experiments.
Single-layer (L20, 71% depth) NLA pair — the original proof of concept, kitft-interface-compatible.
Universal multi-layer NLA trio for Qwen2.5-7B-Instruct (28 layers).
NLA on open-weight models: live demo, corpus, and the Jacobian lens. Repo: github.com/anicka-net/nla-at-home
Qwen coder fine-tunes for backporting CVE fixes.
-
anicka/cve-backport-codegen-v5-qwen25-32b
Text Generation • 33B • Updated • 36 • 2 -
anicka/cve-backport-codegen-v5-qwen3-coder-30b-a3b
Text Generation • Updated • 5 -
anicka/cve-backport-codegen-qwen25-32b-v1
Text Generation • 33B • Updated • 32 • 1 -
anicka/cve-backport-codegen-qwen25-32b
Text Generation • 33B • Updated • 30
Ethics-from-consequences post-training line across model families.
Universal multi-layer NLA pair + AR-native GRPO for Phi-4-mini.
Smallest universal NLA trio (36 layers, one adapter per role). Biggest relative GRPO gain in the series (+31%).
Cross-transliteration Tibetan embedding work.
Qwen coder fine-tunes for backporting CVE fixes.
-
anicka/cve-backport-codegen-v5-qwen25-32b
Text Generation • 33B • Updated • 36 • 2 -
anicka/cve-backport-codegen-v5-qwen3-coder-30b-a3b
Text Generation • Updated • 5 -
anicka/cve-backport-codegen-qwen25-32b-v1
Text Generation • 33B • Updated • 32 • 1 -
anicka/cve-backport-codegen-qwen25-32b
Text Generation • 33B • Updated • 30
Steering vectors and probes from the valence-readout / introspection-frame experiments.
Ethics-from-consequences post-training line across model families.
Single-layer (L20, 71% depth) NLA pair — the original proof of concept, kitft-interface-compatible.
Universal multi-layer NLA pair + AR-native GRPO for Phi-4-mini.
Universal multi-layer NLA trio for Qwen2.5-7B-Instruct (28 layers).
Smallest universal NLA trio (36 layers, one adapter per role). Biggest relative GRPO gain in the series (+31%).
NLA on open-weight models: live demo, corpus, and the Jacobian lens. Repo: github.com/anicka-net/nla-at-home