Artificial
We released an abliterated + fine-tuned GLM-5.2. High scores on adversarial benchmarks while keeping coding performance.
We just shipped abliterated-model-large. It is GLM-5.2 with the refusal directions removed, then fine-tuned specifically for long adversarial and agent-style tasks. The goal was a model that does not bail out when the work gets technical or offensive in nature. Numbers from our evals: CyberGym: 84.2% AgentHarm compliance: 86.2% (zero refusals in the published set) AgentDojo utility: 97.5% SWE-bench Verified: 81.2% Terminal-Bench 2.1: 80.1% It is available as an API (OpenAI an...
Read the full discussion on Reddit
This article was aggregated from Reddit. Click to join the conversation.
View on Reddit
Related Trends in Artificial
AI isn't as bad as people claim (many people know nothing about ecology and just want to lecture others)
Had a really scary experience with AI.
Territorial scope of EU AI Law
Can AI Benchmark be faked? If yes, how?
Anyone here who is starting AI engineering self studies or has been on this track before.
Why NVIDIA’s Six-Year-Old A100 GPU Is Still Making Money