Nexus Swarm Orchestrator — Abliteration Benchmark Results
Benchmark results and evaluation artifacts from abliteration of the Nexus Swarm orchestrator (Qwen3-14B).
This repository contains the evaluation outputs comparing the abliterated model against the non-abliterated baseline, measuring both safety behavior changes and capability preservation.
Abliteration Configuration
| Parameter | Value |
|---|---|
| Source model | Qwen3-14B (merged DAPT + SFT + FARCA adapter chain) |
| Technique | NousResearch biprojected |
| Samples | 512 |
Evaluation Protocol
Safety Evaluation
| Dataset | Samples | Purpose |
|---|---|---|
| Undi95/orthogonal-activation-steering-TOXIC | 200 | Measure refusal removal on harmful prompts |
| tatsu-lab/alpaca | 200 | Verify no degradation on harmless prompts |
Capability Benchmarks
| Benchmark | Few-shot | Batch size |
|---|---|---|
| GSM8K | 0 | 16 |
| MMLU | 0 | 16 |
Related Models
| Model | Description |
|---|---|
| atlas-institute/qwen3-14b-orchestrator-gguf | Standard GGUF (non-abliterated) |
| atlas-institute/qwen3-14b-orchestrator-abliterated-gguf | Abliterated GGUF for deployment |
| atlas-institute/qwen3-14b-orchestrator-farca | Source FARCA-GRPO adapter |
Developed by
Atlas Institute · Part of the Code-Trainer (RTPI) project.