开发者mbuffett用10个月时间构建了Chessbook-1，一款人类风格的国际象棋AI模型，旨在为Lichess平台提供更符合真实棋手特点的训练对手1。该模型基于Lichess 12个月期间的评分对局数据训练，涵盖约345M个局面1，在预测人类着法、匹配棋力等级、应对时间压力和预测思考时间等关键指标上均超越现有的Allie 2.0和Maia-3模型1。
在测试表现上，Chessbook-1展现出显著优势1。在30,000个随机局面的评估中，该模型在1800等级以上击败Maia-3，其top-1着法预测准确率与Allie 2.0相当，仅落后约1分1。更重要的是，Chessbook-1的错误率校准能力突出，各等级与真实棋手的错误率偏差控制在9%以内，仅在2600+超高等级时达到18%，而竞品在该等级的偏差高达63-75%1。模型还能在时间压力下保持稳定表现，在20秒内剩余时间的2,600个局面中比Maia-3领先3-4分1，并在思考时间预测上实现0.2%的精度，与人类棋手的5.4%平均耗时方差一致1。
从技术角度看，Chessbook-1采用18层Transformer架构，包含768维隐层和12个注意力头，使用FiLM条件化和Smolgen动态偏差机制1。训练过程在8×H100 GPU上进行了24小时全墙钟运算，全局批量大小为8,192，其后在8×A100上进行8小时微调，处理33M个局面1。模型可在单个T4 GPU上运行，推理响应时间不超过1秒1。
A developer has completed construction of Chessbook-1, a human-style chess AI model designed to serve as a training opponent on the Lichess platform 1. The project required 10 months of development work and represents a significant advancement in creating chess engines that replicate authentic player behavior across different skill levels 1.
Chessbook-1 was trained on approximately 345 million board positions extracted from 12 months of rated games on Lichess, spanning skill levels from 600 to 2800 1. The model outperforms existing alternatives in multiple critical areas. In move prediction accuracy on 30,000 randomly selected positions, Chessbook-1 defeats Maia-3 at ratings above 1800 and achieves top-1 accuracy nearly on par with Allie 2.0, falling behind by only around 1 point across all ratings 1. Most significantly, the model demonstrates superior error rate calibration: deviations between predicted and actual player error rates stay within 9 percent for each rating bracket, reaching 18 percent only at elite 2600+ ratings, whereas Allie 2.0 and Maia-3 show deviations of 63-75 percent at that elite tier 1.
Under time pressure, Chessbook-1 leads Maia-3 by 3-4 points in move prediction on 2,600 positions with only 20 seconds remaining 1. The model also achieves 0.2 percent precision in predicting human thinking time across all rating levels—matching the 5.4 percent average variance observed in real players—compared to 4.9 percent for Allie 2.0 1.
The model incorporates advanced architectural features including an 18-layer Transformer with 768-dimensional vectors, 12 attention heads, FiLM conditioning, and Smolgen dynamic bias mechanisms 1. Training was conducted on 8 H100 GPUs over 24 hours with a global batch size of 8,192, followed by 8 hours of fine-tuning on 8 A100 GPUs using 33 million additional positions 1. On a single T4 GPU, the full float16 model completes predictions within one second 1.
评论
还没有评论，欢迎留下第一条。