experiment-llm/reports/loss-small.csv at 921069c5db3f27f6906ba8187a29c7b2dbe8b596

Files

T

rzen 921069c5db Add degradation analysis + backing artifacts

- reports/degradation-analysis.md: interpretation of the one-parameter-at-a-time
  ablations (ctx/short/small/data vs v1), grounded in val loss + sample text.
  Key findings: held-out loss tracks quality for generalizing models; different
  degradations give qualitatively different failure text; data-starvation
  overfits (train ppl 1.1 / val ppl 322) with samples that hide the damage.
- reports/compare.md: side-by-side samples across all configs
- reports/loss-{small,short,ctx,data}.csv: variant training curves

2026-07-12 18:28:43 -04:00

1.2 KiB

Raw Blame History

1	iter	train_loss	val_loss	lr
2	500	3.9623	3.9608	0.000600
3	1000	3.4403	3.4320	0.000598
4	1500	3.1455	3.1442	0.000594
5	2000	2.9553	2.9633	0.000589
6	2500	2.8516	2.8464	0.000582
7	3000	2.7778	2.7718	0.000574
8	3500	2.7201	2.7195	0.000564
9	4000	2.6785	2.6591	0.000552
10	4500	2.6386	2.6377	0.000540
11	5000	2.6045	2.5929	0.000525
12	5500	2.5917	2.5906	0.000510
13	6000	2.5724	2.5575	0.000494
14	6500	2.5456	2.5374	0.000476
15	7000	2.5357	2.5335	0.000458
16	7500	2.5279	2.5149	0.000438
17	8000	2.5107	2.4987	0.000418
18	8500	2.5012	2.4953	0.000398
19	9000	2.5002	2.4832	0.000377
20	9500	2.4812	2.4738	0.000356
21	10000	2.4779	2.4697	0.000334
22	10500	2.4721	2.4516	0.000313
23	11000	2.4595	2.4584	0.000292
24	11500	2.4524	2.4410	0.000271
25	12000	2.4597	2.4465	0.000250
26	12500	2.4451	2.4416	0.000230
27	13000	2.4431	2.4348	0.000210
28	13500	2.4292	2.4232	0.000191
29	14000	2.4176	2.4196	0.000173
30	14500	2.4216	2.4238	0.000156
31	15000	2.4285	2.4162	0.000141
32	15500	2.4089	2.4081	0.000126
33	16000	2.4085	2.4228	0.000113
34	16500	2.4114	2.4084	0.000101
35	17000	2.4089	2.4052	0.000090
36	17500	2.4007	2.3989	0.000081
37	18000	2.4007	2.3943	0.000073
38	18500	2.4035	2.3914	0.000068
39	19000	2.4030	2.4014	0.000063
40	19500	2.3995	2.3953	0.000061

1.2 KiB Raw Blame History

1.2 KiB

Raw Blame History