Howdy! We have discovered non deterministic behavior stemming from differing GPU types between provers and verifiers. To offer a redundant verification if hash equivalence fails, a sentence similarity has been added. The similarity threshold is determined based off rank. This is temporary until we can support multiple available GPU types. Tier promotion limits have also been reduced to promote a healthier incentive curve. Provers should expect to see improvements in their scoring.
What's Changed
• Added documentation in the challenge.py file for new functions
• Reduced tier promotion total successes
• Added similarity threshold based off tier
• Increased default sample size
• Added fallback verification with sentence similarity
Patch Notes v1.0.5
Howdy! We have discovered non deterministic behavior stemming from differing GPU types between provers and verifiers. To offer a redundant verification if hash equivalence fails, a sentence similarity has been added. The similarity threshold is determined based off rank. This is temporary until we can support multiple available GPU types. Tier promotion limits have also been reduced to promote a healthier incentive curve. Provers should expect to see improvements in their scoring.
What's Changed
• Added documentation in the challenge.py file for new functions • Reduced tier promotion total successes • Added similarity threshold based off tier • Increased default sample size • Added fallback verification with sentence similarity
Tier Total Successes
Challenger =
100.000-> 4.000 Grandmaster =50,000-> 2,000 Gold =5,000-> 500 # 50 Silver =1,000-> 250Sentence Similarity Thresholds
Challenger = 98% Grandmaster = 95% Gold = 90% Silver = 85% Bronze = 80%