soham

Ordered by Recently added.5 submissions is too few for engagement to mean anything, so this is newest-first.

Everything soham has done here. Publishing an artifact is one way to take part and not the only one: evaluating somebody else's submission, attacking it, and reporting what happened when you used it are the rest, and they are what makes anything here evidence rather than a file.

Submissions5
Labels claimedpro-human, trauma
Kindsdirection
Attacks received0

Namespace

Claimed on by the account custom:huggingface:62cf4580e7f6014c0ea2450f, which called itself sohampadia that day

The claim is bound to the provider's subject id above. Handles are renameable on every provider here, so a claim bound to one would either break on a rename or follow the handle to whoever registers it next. The handle is recorded as of the claim date and is not looked up by.

Evidence

repo

The artifact pinned by soham/trauma@d61-diffmeans-expository-L34 is published at huggingface.co/sohampadia/pro-human, commit 2ff771e4bbdc21a74d37dee111e98be6d71600f5, path vectors/events/trauma_minus_neutral_expository.safetensors. That repository belongs to the account this claim binds, and anyone can check both halves without asking anybody.

recorded

Organisation membership

  • Membership of controlbun confirmed on , with the role admin, by https://huggingface.co/oauth/userinfo.

Each line above is what the provider answered on that date and not a standing fact. People join and leave organisations, so a membership checked once and held forever goes stale silently. Rechecking means the account signs in again and a new observation is written beside this one.

Signing in works and produces the dated reading a claim is made from; binding one to a namespace is still a command run where the corpus is rebuilt, so a claim made today is one somebody handed over rather than one they recorded here. What signing in is for, and what a claim is.

Submissions

ReleasedEvidence
trauma@d61-diffmeans-expository-L34
direction @d61-diffmeans-expository-L34
meta-llama/Llama-3.3-70B-Instructno score, no recipe, never attacked
pro-human@L24
direction @L24
allenai/Olmo-3-1125-32Bno score, recipe, never attacked
pro-human@logistic
direction @logistic
allenai/Olmo-3-1125-32Bno score, recipe, never attacked
pro-human@lda
direction @lda
allenai/Olmo-3-1125-32Bno score, recipe, never attacked
pro-human@meandiff
direction @meandiff
allenai/Olmo-3-1125-32Bno score, recipe, never attacked

Work on other people's submissions

soham has not evaluated, attacked or reported use of anybody else's submission. Nothing follows from that about the quality of their own work. It does mean every number on their submissions was measured by the person who made them.

Eval suites

SuiteJudgeConfound axes checked
confound-audit@v1no judge pinnedapproach, valence, length_independent, approach_probe

The judge model and revision are pinned by the suite's author for their own measurements. A score from a different judge is a different measurement, which is why suites are named rather than pooled.