- So much that i need to take time to fill this in
- Task basked checking to the orchestrator server
- Async DB interactions
- Limiting length of synthetic generation prompts
- Posting more to tauvision again
- Fix for miners who sometimes get partial llm lines streamed
- Pinning Numpy version to fix version issues
- Hotfix to Capacities diminishing every period
- Hotfix to capacities calculation
- Hotfixes to uid period scores
Refactored Weights Setting:
- Made the weights setting process modular.
Fixed Quality Score Finder Bug:
- Corrected the issue in the quality score finder to ensure accurate scoring.
Indexed Database:
- Indexed the database to improve performance and retrieval times.
Increased Synthetic Rate:
- Increased the rate at which synthetic querying happens.
Enhanced Task Selection for Scoring:
- Selected tasks for scoring based on the lack of scores in reward data, so a more accurate picture of a miner is gathered.
Adjusted Period Decay Score:
- Reduced the decay rate of period scores over time.
- Validators will auto adjust to maximize the capacity of a miner on 60-min intervals 🚀
- Miners can chose to rate limit by adjusting their task config files (be careful with your HTTP 429s 🛑)
- Dropped Finetune for now 👋
- Dropped GoAPI SOTA API requirement (savings here) 🫡
- Miners will have massively increased visibility onto their performance through dashboards 🔥
- Validators synthetically score miners on these requests, calculating a
period_score, which represents reliability. - Validators workflow for each task:
- Use the quality scores calculated from the scoring of all tasks for this UID, with a preference for quality scores from this specific task.
- Use a decaying weighted average of previous period scores to calculate a weighted period score.
- Combine the weighted period score with the task quality score and multiply by volume to get the effective volume (capacity).
- Compare the effective capacities with others to get a score for this UID for the task.
- Weight all scores across each task using the subnet-level task weights to derive an overall incentive score.
- Miners can rate limit explicitly to validators without incurring a greater penalty, depending on volume.
- Bittensor upgrade to version 6.9.3.
- Validators to store extensive stats information.
- Reworked weight setting to utilize announced capacities and 'evidence' of those capacities.
- Increased range for speed scores.
- Posted stats for sexy dashboards.
- Small bug fixes for validator proxy.
- Implemented smart load balancing between miners for organic requests using a doubly linked list.
- Cleaned up outdated code.
- Separated validator into specific duties.