註冊並分享邀請連結,可獲得影片播放與邀請獎勵。

George Maloney
@george_onx
Co-founder @ | Creators of GLiNER and GLiDE
加入 November 2024
156 正在關注    1.4K 粉絲
Yesterday we released GLiDE v1.2, now #1# on the Hugging Face Decision Index. We also published how we trained it. The first GLiDE used adaptive reasoning for difficult questions. Strong results but reasoning naturally increased latency. v1.2 has much better performance and latency without any reasoning. -71.5 on the Decision Index -11.3 points ahead of Jev -93ms median leaderboard latency -Better performance than the reasoning version Training GLiDE v1.2 taught us four main lessons: 1. Teach reasoning then distill it into the weights -Reasoning teachers provide soft probability targets -Student answers in one forward pass -Eventually surpassed its teacher at 100x+ lower latency 2. Copy the exam protocol not its answers -Match question decomposition + natural base rates -ACOS jumped from 27.2 → 61.6 -Oversampling alone barely helped 3. Route teachers by skill -Different teachers for different task families -Weights fitted out of fold on held-out rows -1.45 points better than the best single teacher 4. Teach abstention from the model’s own knowledge -Student confidence determines when “I don’t know” is correct -No special head or inference rule -Non-hallucination improved from 16 → ~78 A lot didn’t work too. Short patch fine-tunes hurt hidden performance. Recurrent layers added latency without improving accuracy. FP4 was faster but lost points. See our latest blog for more details. Blog: GLiDE API: Weights:
顯示更多