Unverified50% confidenceBenchmarkExact time
On the SciFact dataset (300 queries, 5,183 scientific abstracts, 768 dimensions), the int8+int4 repack model retained 97.9% of the original float32 version's nDCG@10 (0.836 vs 0.854).
1
Sources
50%
Confidence
Long-term
Relevance
10/13/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedRTX 3060的INT4 Tensor Core理论峰值算力约为101 TOPS，而FP16仅约12.7 TFLOPS，差距达8倍以上61% similarUnverifiedIn the task of completing the last Token of an FP8 quantization function, after DeepSeek-V4-Flash-Base adds equal-sign decorations before the code, when the number of equal signs modulo 4 is 0 or 1 the error rate is as high as 71.3%, and when the remainder is 2 or 3 the accuracy returns to 91.5%.61% similarUnverifiedAdding 3 neutral paraphrased versions per cell beyond the original prompt, 0 out of 112 configurations qualified (0/112).60% similarUnverifiedThe self-developed W4 kernel's perplexity (probe ppl) is 15.76, better than bitsandbytes NF4's 18.98.59% similarUnverifiedUnder the optimal configuration, the EduBehaviors framework achieves a macro F1 score of 0.673 and a Cohen's kappa of 0.688.59% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1014140API
curl https://kongchang.com/api/v1/knowledge/claims/1014140MCP
get_claim(id=1014140)