Unverified50% confidenceBenchmarkExact time
After fine-tuning the 406M model specifically for the 2,000-word retrieved segment input format, the previously lost 6.30 points were fully recovered.
1
Sources
50%
Confidence
Long-term
Relevance
9/23/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedAblation analysis shows that fine-tuning on the segment paradigm yields a +5.29-point improvement on the test set, while replacing the first 4,500 transcript words with 2,000 retrieved words yields only a +1.55-point improvement.68% similarUnverifiedOLLAMA_KV_CACHE_TYPE=q8_0 occupies approximately only half the memory of the f16 default, with minimal quality loss64% similarVerified4-bit quantization reduces model size to 1/8 of the original at the cost of approximately 1–3% accuracy loss.63% similarVerifiedINT8 quantization loses approximately 0.5% accuracy while halving VRAM usage; INT4 quantization compresses VRAM to 1/4 of the original with an accuracy loss of approximately 1–3%.63% similarUnverified8-bit quantization nearly halves memory usage with minimal quality loss and strong instruction-following capability.63% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/942873API
curl https://kongchang.com/api/v1/knowledge/claims/942873MCP
get_claim(id=942873)