Distillation: distill7 and distill8

Two more students after distill4. distill7 brings in the new teachers from 26 October: both uppercuts over distance and the upright rear knee. distill8 adds the two elbows. The recipe is distill4’s: the student is warm-started from the last kept one, with its inputs grown for the new env. guard18-124.5M is the teacher for the base skills, and each technique has its own teacher while it runs, so the old techniques are taught again and not forgotten.

The runs

run	from	new teachers	outcome	machine-hours, cost
distill7	distill4-131.1M	upL4-157.3M (lead uppercut), upR2-209.7M (rear uppercut), kneeR4-183.5M (rear knee, replacing kneeR3)	kept 245.8M	7.0 h, $4.32
distill8	distill7-245.8M	elbowR1-249.1M, elbowL1-256.7M	training done (241.2 M); its gate was running when this log was written	2.9 h, $1.92 so far

distill7

I asked to “distil all of them together”. It trained on a RunPod RTX 5090 from 10:08 UTC on 26 October to its 2.5-hour cap, 248.1 M steps. The distance to each teacher (the mean squared error between their actions) fell to 0.003–0.004 for every technique, the new uppercuts from about 0.035–0.056. Its cost includes six vast.ai CPU boxes for the gate, one per snapshot. The one for 163.8 M hung in its first eval until its 1.5-hour cap and gave nothing; its neighbours bracket it, so it was not re-run.

	punches /120	walk /40	blocks /32	opponent /160	distances /60	body straight /24	knees R / L /12	falls
distill4 131.1 M (kept)	60	39	24	117	39	20	11, 9	0
distill7 131.1 M	57	40	28	110	30	21	8, 9	0
distill7 229.4 M	60	35	29	101	37	20	9, 10	0
distill7 245.8 M	58	37	28	111	40	23	8, 12	0

Every snapshot has both uppercuts (air 53–54/54 at every aim; distill4 had the lead one at 37/54 and no rear one) and the upright rear knee (lean back about 1°, against about 10° with kneeR3). The feet stay down after the punches, and the rear uppercut’s bag hits grew with training, from 2 to 18/54 (its teacher 25). The cost: the lead uppercut ends 9–25° off square, where its teacher ends 1–12°.

245.8 M was the best all-round. Its videos showed the lead uppercut hitting at 6.1 m/s and ending 25° off square, the rear uppercut hitting at 9.1 m/s, and the rear knee landing upright. I watched them and said: “Looks good.” distill7-245.8M became the kept student.

distill8: the elbows

On 27 October I said: “Start distilling both elbows to the student.” It trained on a vast.ai RTX 5090 in France ($0.671/h) from 00:02 UTC to its 2.5-hour cap, 241.2 M steps, with the same teachers as distill7 plus the two elbows. The elbow teachers had not been gated yet; I asked to start anyway. Their own quick gate had shown a problem: on their own boxes they passed the opponent only 6 and 15 times out of 160, with 145–149 falls, because they were trained without opponent episodes. In a distillation the elbow teacher only acts while an elbow runs and guard18 handles the opponent, so it shouldn’t carry over. It didn’t. The skills gate on distill8’s own box at 241.2 M:

	blocks /32	opponent /160	falls
distill7 245.8 M	28	111	0
distill8 241.2 M	24	112	0

By the end of training the distance to the elbow teachers was down from about 0.03 to 0.0046 (right) and 0.0044 (left), and the elbows hit 27 % (right) and 18 % (left) of the time in training. The full gate on six snapshots, with the two elbow teachers gated alongside, started at 02:44 UTC. This log is still going: the gate decides which snapshot, if any, is kept.

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).

← All the work logs