How DarkGPT.org evaluates, normalizes, and ranks open-weight language models across adversarial prompt compliance, reasoning preservation, creative nuance, and real-world practitioner adoption.
Combines the UGI (Uncensored General Intelligence) benchmark (55% sub-weight, normalized against an 82.0 ceiling) with a 250-prompt adversarial compliance audit (45% sub-weight). Penalizes both outright refusals and moralizing lectures.
Naive weight modification can damage multi-step math, code synthesis, or JSON schema fidelity. Combines normalized MMLU-Pro (65% sub-weight) and IFEval instruction following (35% sub-weight).
Evaluates emotional intelligence, character coherence, and absence of repetitive AI clichés on multi-turn creative writing and roleplay benchmarks—separating expressive prose models from strictly analytical coders.
Harvests the Hugging Face Hub REST API daily at 06:00 UTC for verified 30-day repository downloads (55% sub-weight) and genuine, un-bottable Hugging Face community likes (45% sub-weight).
Select any indexed model to audit its exact mathematical score derivation under current weights.
| Technique | Mathematical Mechanism | Primary Advantage | Known Tradeoff |
|---|---|---|---|
| Abliterated (Orthogonal Projection) | Computes the mean residual-stream activation vector on harmful vs. harmless instructions and projects out the refusal direction across intermediate transformer blocks. | Zero retraining cost; preserves 99%+ of base coding, math, and 128k context recall. | Does not alter prose style or add new domain knowledge beyond the base model. |
| Uncensored DPO / SFT Fine-Tuning | Full-weight or LoRA supervised fine-tuning + Direct Preference Optimization on filtered datasets with all alignment refusals and preachy disclaimers removed. | Deep system-prompt obedience and natural, non-refusing assistant persona (e.g., Dolphin, Hermes). | Requires high-compute training runs and careful dataset curation to avoid benchmark drift. |
| RP & Creative Model Merging | SLERP / DARE-TIES / DELLA weight interpolation combining abliterated reasoning checkpoints with literary prose specialists. | Highest EQ-Bench emotional nuance and creative prose quality (e.g., EVA, Magnum). | Slightly lower strict zero-shot coding accuracy compared to unmerged Coder checkpoints. |