SolsticeBenchgen4 (Solstice self-play)
| Progress | |
|---|---|
| Epoch | — |
| This epoch | |
| Overall | |
| Running loss | — |
| Started | — |
| Elapsed | — |
| Remaining | — |
| Validation | |
| Best | — |
| Latest | — |
| Accuracy | — |
| Run | |
| Host | chess-engine-test-top-level |
| arch | fly connectome |
| batch_size | 512 |
| cache | gen4 (Solstice self-play) |
| epoch_unit | one validation interval = 2,000 steps = 1,024,000 positions |
| lr | 0.003 |
| min_synapses | 5 |
| parameters | 2,700,513 synapse strengths |
| policy_head | lc0-style move policy from the descending neurons (weight 0.1); accuracy column = how often its top move matches Solstice's |
| positions_per_second | 674 |
| positions_seen | 45,568,000 |
| recurrent_steps | 6 |
| synapses_trained | True |
| wdl_lambda | 0.75 |
Loss by epoch
Validation · Train
Train loss
Epochs
| Epoch | Train | Val | Acc | LR | Time |
|---|