Qwen 3.8 27B shows different rejection traits than Gemma4; plan to generate 10k prompts for analysis
max_paperclips · x · 2026-08-15
Researchers observed significant differences in rejection characteristics between Qwen 3.8 27B and Gemma4 31B. To extract the "Ablirated" vector for the 27B model, the team plans to recreate the prompt set and generate approximately 10,000 prompts in their home lab, specifically focusing on those rejected by the Qwen model.
More from Models
- Zhipu's GLM 5.3 Gains on Same Base Model, Highlighting Post-Training Power — oran_ge · 2026-08-15
- DeepSeek Vision offers the only truly instant free vision experience right now — teortaxesTex · 2026-08-15
- Gemini 3.7 Flash spotted, praised for performance potentially boosted by RL — AI_Andrew · 2026-08-15
- DeepSeek V4 on Cloudflare Workers AI: First 1M Token Context Models — ritakozlov · 2026-08-15
- Running LTX-2.5 on 8GB VRAM + 16GB RAM: A user's experience — Rique_Belt · 2026-08-15
- Dev Reports Better Performance with Qwen 3.8 Reasoning Disabled — Jordanthecomeback · 2026-08-15