Unverified50% confidenceFactExact time
The fixed test conditions included: base model unsloth/Llama-3.2-3B-Instruct-bnb-4bit, 1000 samples, 2 epochs, 250 steps, batch size 2, gradient accumulation 4, effective batch size 8, and about 24.3M trainable parameters (0.75% of total parameters).
1
Sources
50%
Confidence
Medium-term (~90 days)
Relevance
11/10/2026
First Seen
Valid until: 2/8/2027
Sources
Related Entities
Related Claims
UnverifiedThe limitations of this test include: limited task scale (50 tasks, 3 trials per prompt), style versions written by a single author, length confounding between the florid and normal versions, testing only English and local models below 14B, and the perplexity check covering only three smaller Qwen models and using full-precision weights while actual inference used 4-bit quantization.72% similarUnverified叠加cv=5的5折交叉验证后,144种参数组合需训练144×5=720次模型71% similarUnverifiedOn a stratified 1,000-item test subset, the strongest LLM prompting configuration achieved an accuracy of 0.477, while the fine-tuned DistilBERT reached 0.879.70% similarUnverifiedA 360M-parameter code model trained by the trifero.ai lab under Strife Technologies initially reported 26.8% pass@1 on HumanEval, and after removing instruction data overlapping with the benchmark, the true number was 22.6%.70% similarUnverifiedSuggested default retry parameters: base of 200ms, cap of 30000ms, and max of 5 attempts (plus the initial call for 6 total).69% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1157267API
curl https://kongchang.com/api/v1/knowledge/claims/1157267MCP
get_claim(id=1157267)