OpenTransformer commited on
Commit
70bd139
·
verified ·
1 Parent(s): 0cebf91

Upload inference_results/step15015978_20260319_2253.md with huggingface_hub

Browse files
inference_results/step15015978_20260319_2253.md ADDED
@@ -0,0 +1,26 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # Inference Results - Step 15015978 (20260319_2253)
2
+
3
+ ## AR Mode
4
+ **Prompt:** "The principles of fluid mechanics explain how submarines navigate"
5
+ **Output:** matterHowanavian-receptorfunctionssurfacesthecellsthatinventsthecellsofMC1,5and7,respectively.Melanocortis(12)isthenembeddedreceptorsignalingpathwayofMC1,5,and7,respectively.NucleicAcids(13)aretheresultofMC1and5,whicharetheresultofMC
6
+
7
+ **Notes:** No spaces between words. Topic drifts from submarines to melanocortin receptors/nucleic acids. Uses scientific terminology (MC1, MC5, MC7) and citation-style references. Semi-coherent English structure but nonsensical content relative to prompt. 1.2 tok/s, 65.71s.
8
+
9
+ ## SAT Fixed Mode
10
+ **Prompt:** "how radio waves carry information through the atmosphere"
11
+ **Output:** stareil1GWinartRehatiansinthetoouttheatesMumieestHlandand.Sursser|is:PontDwasTheasakFJortrefanseecforar".agofhedasszyusotismitopiritsacInireK\xef\xbf\xbdemutm.P.comEoftwiththoricim
12
+
13
+ **Notes:** Garbled subword fragments. Not coherent English. 2.4 tok/s, 33.00s.
14
+
15
+ ## SAT Variable Mode
16
+ **Prompt:** "the mathematics of probability theory and its origins in gambling"
17
+ **Output:** im:intheutReareouteW.SMsbytoiniliepeswasthandBst|asyss's.HJarianonturHenlandGist"Arfalsestusag.PatPisassureKFec4oftakirhedwith;whe.comismTheortireerastit1whichhatEthator
18
+
19
+ **Notes:** Garbled subword fragments. Not coherent English. 2.4 tok/s, 33.48s.
20
+
21
+ ## Notes
22
+ - Checkpoint: pretrain_step15015978.pt (~13.95B/35B tokens, 39.8%)
23
+ - Inference run on CPU (GPU busy with training)
24
+ - max_new_tokens: 80
25
+ - AR mode produces no-space English with scientific vocabulary and topic drift — consistent with previous observations
26
+ - SAT modes (both fixed and variable) still produce garbled subword fragments at 39.8% training — no improvement from previous check