OpenTransformer commited on
Commit
22c1461
·
verified ·
1 Parent(s): 6ea9d60

Upload inference_results/step14854370_20260319_0524.md with huggingface_hub

Browse files
inference_results/step14854370_20260319_0524.md ADDED
@@ -0,0 +1,41 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # Inference Results — step 14854370
2
+ **Date**: 2026-03-19 05:24 UTC
3
+ **Checkpoint**: pretrain_step14854370.pt
4
+ **Training progress**: 39.1% (~13.69B/35B tokens)
5
+ **Model**: 698.39M params, d=1024, 24 layers, 16 heads, rank 128, 2x expansion
6
+ **Device**: CPU (GPU occupied by training)
7
+
8
+ ---
9
+
10
+ ## 1. AR (Autoregressive)
11
+ **Prompt**: "The development of calculus independently by both Newton and Leibniz demonstrated"
12
+
13
+ **Output**: itslowlyadvancedSt.Louis'spopulationonafirmerchantarea.TheDukeMuseumhasbeenbuilttopaythe.4millioninterestthatfirmlyendedthepracticeofthepresident'smeeting.TheDukeMuseumacknowledgedthatthenewFoxNews"isthetimeforthelastfewmonths,butitishardforfuturetalents."
14
+
15
+ **Notes**: No spaces between tokens (known issue). Semi-coherent English sentences. Hallucinates Duke Museum/Fox News. No repetitive loops. Some grammatical structure present.
16
+
17
+ ---
18
+
19
+ ## 2. SAT Fixed
20
+ **Prompt**: "how the process of continental weathering shapes landscapes over geological time"
21
+
22
+ **Output**: ),NastestOittemup'usrehehedhatBorhadfCieasVthat1910Sthesyoriginalright.thethePbthetheWtheayserthecthISBNagainstlist8utfrom20F]et:atingidDed28e.[9R0p.Sthe1st12LeíoftNewafter4
23
+
24
+ **Notes**: Garbled subword fragments. SAT fixed still broken at 39.1%.
25
+
26
+ ---
27
+
28
+ ## 3. SAT Variable
29
+ **Prompt**: "the engineering challenges of building space elevators connecting Earth to orbit"
30
+
31
+ **Output**: emOop)intheitassreR,tCest10inersac19rwith,andalsTheardusans]".20edersartastabliode).Le.comamaricentythasout,Sallish25(.SlareDaySanI12radationynakbandzwardryrestepMay124ofunW8AnA
32
+
33
+ **Notes**: Garbled subword fragments. SAT variable still broken at 39.1%.
34
+
35
+ ---
36
+
37
+ ## Summary
38
+ - AR: Producing coherent English sentences (no spaces). Quality similar to previous checks.
39
+ - SAT (both modes): Still producing garbled output at 39.1% training progress.
40
+ - Loss range: 4.2-13.5 (typical, occasional spikes)
41
+ - GPU throughput: ~19.4k tok/s