Work log · Teacher · Pull counter
Work log for Shadow clones: four Muay Thai techniques learnt in parallel
Teacher: the pull counter
Training cost about $3.454 vast.ai boxes that never loaded, 1 RunPod RTX 5090 for training and 3 for evals, 24–25 October 2026 · 1 training runs, 8 machines, about 3.1 machine-hours
The pull counter, from my motion capture (Muay Thai 02, take_01, 16.90–18.45 s): lean back away from a punch with the right fist wound back, then drive forward into a rear straight to the head, and back to the guard. Its own environment (opponent_pull), warm from distill1-282.5M, which held every other skill.
Four vast.ai boxes in a row never loaded: two had no SSH 4 minutes after starting, and two were still loading after 10 minutes. Together they cost $0.49. The run went to RunPod instead and trained to its 2 h cap, 212 M steps ($2.48).
| snapshot | air hits (median miss, speed) | bag hits | falls |
|---|---|---|---|
| 104.9 M | 3/8 (9.7 cm, 2.5 m/s) | 8/8 | 0 |
| 157.3 M | 4/8 (9.0 cm, 2.8 m/s) | 8/8 | 0 |
| 183.5 M | 8/8 (6.6 cm, 3.5 m/s) | 8/8 | 0 |
| 209.7 M | 8/8 (4.6 cm, 4.7 m/s) | 8/8 (2.5 cm, 7.1 m/s) | 0 |
Kept: pull1-209.7M. In the side view at quarter speed the trunk tilts back with the right fist wound back, then drives forward into the straight: the lean is there, not a plain cross.
Evaluating it took three RunPod pods ($0.49). The first two failed on our own code: the env’s tables were built in daemonic worker processes, and ffmpeg read a Windows D:\ output path as a protocol off Windows. Both are fixed.
Distilled in distill6, the student learnt it at the bag (6/6) but not in the air (1–2/6, about 10 cm short). See the distillation work log.
Every run
From the run records in kakuto-ml: each run's machines, their billed hours and cost (EC2 runs are estimates from the run logs).
| Run | Outcome | Machines | Hours | Cost |
|---|---|---|---|---|
pull1-2026102424 October · pull counter · kept pull1-209.7M | Kept | 8 | 3.1 | $3.45 |
| 1 training runs | 8 | 3.1 | $3.45 |
← All the work logs for “Shadow clones: four Muay Thai techniques learnt in parallel”