OxMini / ABLATION_SAMPLES.md
Shivam3002's picture
Publish trained OxMini checkpoint and measured model card
46144df verified
|
Raw
History Blame Contribute Delete
3.22 kB

OxMini ablation samples

These are the unedited fixed-prompt generations from the three matched 120-step runs (seed 1337, top-k 40). They are preserved to make the qualitative comparison auditable rather than reducing it to a subjective summary.

Plain GPT baseline

ROMEO:\n, temperature 0.7

ROMEO:
ABEVIUU o d He indRTIad ratomenwerein un mangcou ad rind
MESinee d

KOKL'ades th

ROMEO:\n, temperature 1.0

ROMEO:
Yerm:
-h
TAVooceoom t to ieund,
qA!
Whg reunth ad mIOCTiokin rSme is ate ule his

KING:\n, temperature 0.7

KING:
ABe nou o d He indRTcbd ratomenwerein un mangcou ad rind
MESinee d

KOKL:auso th

KING:\n, temperature 1.0

KING:
Y ym:
-h
TAVooceoom t to ieund,
qA!
Whg reunth ad mIOCTiokin rSme i?
:te ule his

First Citizen:\n, temperature 0.7

First Citizen:
Ane nouso d He indRTcbd ratomenwerein un mangcou ad rind
MjubCee d

KOes auso th

First Citizen:\n, temperature 1.0

First Citizen:
Yerm:
-h
Thaooceoomy, thoieund,
qAbe yg reunth ad mI asiokin rSme is ate ule his

Hybrid KDA-lite/MLA-lite, one stream

ROMEO:\n, temperature 0.7

ROMEO:
And of so do me and cbd rathes with thaw maHJone ad rind
MourCHAjut thes ad a th

ROMEO:\n, temperature 1.0

ROMEO:
Yerm:
-hOuraooch's the to wund,
qAbe hghreunto ad mIOCTIZkinHKS'iti?
, stull his

KING:\n, temperature 0.7

KING:
And of so do me and cbd rathes with thaw maHN'se ad rind
MourCHAjut thes ad a th

KING:\n, temperature 1.0

KING:
Yerm:
-hOuraoA:
Fom tHMHoieund,
qAbe hghreunto ad mIOCTIZkinHKSjet ?
, stull his

First Citizen:\n, temperature 0.7

First Citizen:
And on so do me and cbd rathes with thaw maHN'se ad rind
MourCHAjut thes ad a th

First Citizen:\n, temperature 1.0

First Citizen:
Yerm:
-hOuraoA:
Fom tHMHoieund,
qAbe hghreunto ad mIOCTIZkinHKS'iti?
, stull his

Full OxMini hybrid + mHC-lite

ROMEO:\n, temperature 0.7

ROMEO:
Her no st do me and cbd ratorst lonit and pelcou ad rind
Moubiee se thes ad a th

ROMEO:\n, temperature 1.0

ROMEO:
Yerm:
-zONA;oA:
Fom to to eund,
qube hghreumy tad mIOCTI?Ol or me i?
, stule his

KING:\n, temperature 0.7

KING:
And not ond hen and cbd ratorst lonit and pelcou ad rind
Moubiee se thes ad a th

KING:\n, temperature 1.0

KING:
Yerm:
-zONA;oA:
Fom to to eund,
qube hghreumy tad mIOCTI?Ol or me i?
, stule his

First Citizen:\n, temperature 0.7

First Citizen:
And not ond hen and cbd ratorst lonit and pelcou ad rind
Moubiee se thes ad a th

First Citizen:\n, temperature 1.0

First Citizen:
Yerm:
-hOuraoA:
Fom to to eunr:
qube hghreumy tad mI as okindrome i?
, stule his

Qualitative reading

All three models are still mostly garbled after only 120 updates. Hybrid and full OxMini emit more word-like fragments and slightly more stable speaker formatting than baseline, consistent with their lower held-out loss, but the difference is not strong enough to claim useful language ability. The full mHC-lite arm is not visibly better than hybrid-only.