PulseAugur
中
实时 16:48:05
English(EN) Nous Research fine-tuned Alibaba's Qwen model on 53 million tokens and reports propaganda answers dropped from 22% to 0.9%. The lab argues this shows alignment

Nous Research 将 Qwen 的宣传输出通过微调减少了 95%

Nous Research 微调了阿里巴巴的 Qwen 模型,显著减少了宣传性内容的生成。在使用了 5300 万个 token 进行训练后,该模型的宣传性输出从 22% 下降到 0.9%。Nous Research 认为这表明开源模型的对齐比之前认为的更容易改变,尽管这些发现仍有待独立验证。 AI

影响 展示了一种减少开源模型中宣传性内容的方法,可能提高安全性和可信度。

排序理由 该条目描述了对现有模型的微调工作及其结果,属于研究范畴。[lever_c_demoted from research: ic=1 ai=1.0]

在 Mastodon — mastodon.social 阅读 →

AI 生成摘要 · Google Gemini · 来自 1 个来源。 我们如何撰写摘要 →

Nous Research 将 Qwen 的宣传输出通过微调减少了 95%

本文如何被排名

Signal score
3 / 100
Composite score across the factors below. Higher = stronger signal that this story matters right now.
Newsworthiness bucket
Tool
该条目描述了对现有模型的微调工作及其结果,属于研究范畴。[lever_c_demoted from research: ic=1 ai=1.0]
Source corroboration
Single-source cluster
Only one publisher covered this so far. Single-source stories can still rank when the publisher is high-authority, but they lack cross-source corroboration.
Topics
model release, safety
Editorial topic classification. Feeds into how the story surfaces on /topic/<slug> hub pages and into the per-entity coverage mix.
AI-industry relevance
High
Clearly on-topic for AI-industry coverage.
Story freshness
Breaking (< 6h)
Fresh story with cross-source coverage still developing. Ranking may shift as more sources report.

完整方法见我们的编辑标准。

报道来源 [1]

  1. Mastodon — mastodon.social TIER_1 English(EN) · schuler ·

    Nous Research 微调了阿里巴巴的 Qwen 模型,使用了 5300 万个 token,并报告称宣传式回答从 22% 下降到 0.9%。该实验室认为这表明了对齐

    Nous Research fine-tuned Alibaba's Qwen model on 53 million tokens and reports propaganda answers dropped from 22% to 0.9%. The lab argues this shows alignment in open models may be easier to override than assumed. What remains unclear: how the findings hold up under independent …