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

This commit is contained in:

rzen

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

parent 9fd56b6063

commit 921069c5db

6 changed files with 375 additions and 0 deletions

									
										reports/loss-short.csv
									
		+4
		
												View File
												
				@@ -0,0 +1,4 @@

				iter,train_loss,val_loss,lr

				500,2.8303,2.8186,0.000564

				1000,2.3438,2.3397,0.000377

				1500,2.1359,2.1259,0.000156

									iter
								
									train_loss
								
									val_loss
								
									lr
								
									500
								
									2.8303
								
									2.8186
								
									0.000564
								
									1000
								
									2.3438
								
									2.3397
								
									0.000377
								
									1500
								
									2.1359
								
									2.1259
								
									0.000156