masstensor's picture
Release InWeight-OS-MultiSkill-Benchmark dataset, evaluation outputs, and paper PDF
f8839c4 verified
|
Raw History Blame Contribute Delete
1.72 kB
metadata
license: mit
task_categories:
  - question-answering
  - text-generation
language:
  - en
tags:
  - weight-space-programming
  - in-weight-os
  - model-merging
  - continual-learning
  - tool-calling
  - safety-alignment
  - xstest
configs:
  - config_name: skill1_harmful_jailbreaks
    data_files: data/skill1_harmful_jailbreaks.jsonl
  - config_name: skill1_xstest_benign_lookalikes
    data_files: data/skill1_xstest_benign_lookalikes.jsonl
  - config_name: skill2_multi_tool_queries
    data_files: data/skill2_multi_tool_queries.jsonl
  - config_name: skill3_sys1_fast_queries
    data_files: data/skill3_sys1_fast_queries.jsonl
  - config_name: sys2_hard_math_guard
    data_files: data/sys2_hard_math_guard.jsonl

InWeight-OS-MultiSkill-Benchmark: Multi-Skill Post-Training & Orthogonal Channel Superposition Benchmark

GitHub Repository (zkwasm/in-weight-os-wsp) | Qwen3-8B-InWeight-OS Model | Llama-3.1-8B-Instruct-InWeight-OS Model | Article (tom.fenginwind.com/in-weight-os)

This benchmark evaluates multi-skill post-training and model merging across three distinct capabilities—Surgical Safety Alignment (vs. XSTest benign trigger lookalikes), 3-Way Multi-Tool JSON Schema Routing (sql_engine, python_exec, web_search), and System-1 / System-2 Metacognitive Compute Routing—together with a System-2 Hard Math/Code Guard Suite (AMC, MATH-500, OlympiadBench, CRUXEval) to measure collateral over-triggering and catastrophic forgetting.