Upload inference_logs/infer_20260317_015107_step14212405.md with huggingface_hub
Browse files
inference_logs/infer_20260317_015107_step14212405.md
ADDED
|
@@ -0,0 +1,24 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# Inference Test — step14212405 (2026-03-17 ~01:47 UTC)
|
| 2 |
+
## Checkpoint: pretrain_step14212405.pt
|
| 3 |
+
## Training progress: ~12.91B/35B tokens (36.9%)
|
| 4 |
+
|
| 5 |
+
### AR Test
|
| 6 |
+
- Prompt: 'The principles of fluid dynamics govern how liquids and gases'
|
| 7 |
+
- Output: arangeofspecificaspects,suchas"variousformsofmediatoactivatingoutilyministerialseats,fromadverselyregressionofthepersonalityperiodforthepersonalityconsultants,whoimmigratedtoothercountriesatthecoreofhispeacewiththeratificatemodern-dayuseofitsownprimarypromptes,whichwasassessedby
|
| 8 |
+
- Notes: No spaces, severe topic drift into politics/personality. Still garbled.
|
| 9 |
+
|
| 10 |
+
### SAT Fixed Test
|
| 11 |
+
- Prompt: 'the process of stellar nucleosynthesis creates heavier elements inside stars'
|
| 12 |
+
- Output: eciaade.PreatberCh("neletusdelstemerachseSrabollibautup:gur,spGolRterasallepwefofofafacofofarerenhimmeub.LarLserod;inouNcomde:ifiilel's202of:ortZn.SitIGuline
|
| 13 |
+
- Notes: Garbled subword fragments. SAT still not producing coherent text.
|
| 14 |
+
|
| 15 |
+
### SAT Variable Test
|
| 16 |
+
- Prompt: 'the evolution of writing systems from cuneiform to modern alphabets'
|
| 17 |
+
- Output: atuslesillineWachyletstberSjolledecrellajaualeRlificzballiv00roGiclbant'KquienCZichtemBunLesehrthatFranteranverasderovensiaofite'epifInapp:of.Dofobofominassgitdecheapol
|
| 18 |
+
- Notes: Garbled subword fragments. SAT variable still not producing coherent text.
|
| 19 |
+
|
| 20 |
+
### Summary
|
| 21 |
+
- AR: Produces connected words but no spaces, heavy topic drift (prompt about fluid dynamics → politics)
|
| 22 |
+
- SAT fixed: Still garbled subword fragments at 36.9%
|
| 23 |
+
- SAT variable: Still garbled subword fragments at 36.9%
|
| 24 |
+
- Slight improvement in AR coherence compared to earlier checks, but still fundamentally broken output
|