Unverified50% confidenceOpinionExact time
Open-source model safety should have safety properties embedded at the training stage, and relying solely on post-training RLHF alignment is insufficient.
1
Sources
50%
Confidence
Long-term
Relevance
9/17/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedThe safety alignment in the original training of open-source models can be removed through a small amount of fine-tuning.79% similarUnverifiedRLHF和DPO等对齐技术的效果高度依赖预训练基础质量，预训练阶段未习得的能力几乎无法通过后续对齐补救75% similarUnverifiedOpen-source models may have biases and risks embedded from training data, only the attribution of responsibility for those risks becomes more diffuse74% similarVerifiedOpenAI states that safety has been the primary consideration from the first day of model training, and the model is guided to proactively steer conversations away from risky topics.73% similarUnverifiedModel-layer security constraints include alignment training (RLHF, Constitutional AI), system prompt security instructions, output filters, and guardrail models such as Llama Guard and NeMo Guardrails.72% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/930050API
curl https://kongchang.com/api/v1/knowledge/claims/930050MCP
get_claim(id=930050)