研究人員使用錯誤中學習的人類學習法來訓練AI模型
· 2023-11-06

研究人員提出從錯誤中學習(Learning from Mistake)的LeMA訓練法,以提升開源LLM在數學解題等推理任務上的效能

_arXiv:2310.20689 [cs.CL]

研究人員以2種問題資料集(GSM8K及MATH)實驗LeMa方法對5個開源LLM的效果,並比較只以CoT資料集來微調LLM的效果。結果顯示,以LLaMA-2-70B為例,它在兩種方法下,在GSM8K的準確率分別為83.5%及81.4%,在MATH則分別為25.0%及23.6%。此外,他們也實驗了WizardMath及MetaMath二種專門領域LLM的準確率,在GSM8K資料集測試中,獲致84.2%及85.4% pass@1 準確率,而MATH資料集則達27.1%及26.9%,這個成績超越非執行(non-execution)開源模型在同樣任務中的表現。

此外,他們發現,在同樣資料量的訓練集下,LeMA方法也比純CoT微調來得好。此外,整合CoT資料及修正資料,微調效果更優於單一的微調結果。

研究人員已將LeMA的程式碼、模型、資料公開在GitHub上。

Popular articles
1spin4win releases unique slot Don Catleone Hold and Win featuring gangster cats
Online Game
New Jersey July Gambling Revenue Hits $606M, Sweeps Casinos Banned
Regulation
Indiana online casino bill stalls in House committee
Regulation
Super PAC Raises $48 Million: Sports Betting Forces Ramp Up Political Push
Regulation
JILI Partners with Cricket Legend AB de Villiers (ABD) to Launch Exclusive Branded Game Series 100% 11
Sports Game
SBC Summit Canada to Make Player Safety a Key Pillar of 2026 Agenda
Marketing
GAT CDMX 2025 Institutional Academy: Leaders and Experts Analyze the Present and Future of the Gaming Industry in Mexico and Lat
Sports Game
Brazil Proposes Raising Gambling Tax Rate to 24%, With Revenue Allocated to Social Security and Healthcare
Regulation
Kazakhstan plans to penalise online casino promotions
Regulation
Institutional Academy that exceeded expectations marked the opening of GAT CDMX
Online Game
Vietnam’s Controlled Gaming Shift Gains Ground, But Domestic Demand Still Lags
Southeast Asia
GGC Awards 2026 Shines in Colombo: Honoring Leaders and Innovators in the iGaming Industry
HUIDU Focus
Across 6 Cities: HUIDU Invites You to 8 World Cup Parties Redefining High-Value Social Networking
HUIDU Focus
Vietnam's tightening online gaming policy creates new market opportunities
Southeast Asia
Are you ready to maximize your earnings? Try ProPush.me Constructor!
Marketing
Home
Game
Cooperation
Find
My