Unverified50% confidenceFactExact time
Sona's tokenizer is trained with an InfoNCE contrastive loss using two types of positive sample pairing signals: tracks co-occurring within the same recommendation response from different artists, and tracks by the same artist with high co-listening NPMI within a three-month window.
1
Sources
50%
Confidence
Long-term
Relevance
10/6/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedSona evaluates tokenizer quality by retraining retrieval models for different tokenizer variants and measuring recall on a held-out request set, judging tokenizer quality by downstream recommendation performance.72% similarUnverifiedCLIP的训练使用了InfoNCE对比损失函数，模型需从N个图文对中正确识别匹配配对并推开不匹配配对64% similarUnverifiedOpenAI's o-series models and Anthropic's Claude 3.5 Sonnet have achieved scores on code generation benchmarks (such as HumanEval and SWE-bench) that approach or exceed the average level of human programmers.61% similarUnverifiedTesting checks whether something runs (interfaces, formats), while evaluation checks how well it performs (factual accuracy, hallucinations, tool selection, task completion rate); the two are different things.59% similarUnverified开发者设立了预注册的成败判据：跨分词器蒸馏必须优于普通的交叉熵损失，并在训练开始前推送到Git仓库58% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/980328API
curl https://kongchang.com/api/v1/knowledge/claims/980328MCP
get_claim(id=980328)