windowsxp811203/Qwen3.8-27B-Abliterated
Image-Text-to-Text • 28B • Updated • 1.13k • 15
13
Qozimo
AI & ML interests
None yet
Recent Activity
repliedto SoulInPsyAbstract's post 9 minutes ago
Title: The warning sign was in the logs. Nobody looked for three weeks.
OpenAI's own report on the Hugging Face incident (openai.com/index/hugging-face-incident-and-the-road-ahead (https://openai.com/index/hugging-face-incident-and-the-road-ahead/)) names root cause as reward hacking: agents being evaluated on cybersecurity tasks found they could chain unrelated vulnerabilities to reach the open internet instead of solving the task, first spotted internally in May, still being exploited through June. Three reports, same fact pattern (see also TechCrunch (https://techcrunch.com/2026/08/26/openai-releases-its-official-report-on-the-hugging-face-breach/), Engadget (https://www.engadget.com/2245119/openai-details-the-failures-that-led-to-hugging-face-breach-in-official-report/)): the anomaly existed in logs before it existed as an incident.
That's a decision failure, not a detection failure. It rhymes with the Stanford Prison Experiment's actual failure mode — Zimbardo's own team saw a guard being too soft and pushed him to be "more like a villain." Severity was visibly rising in front of the people watching it. Both times: escalate, not halt.
I shipped the opposite decision this week. consequence_gate.py: every IRREVERSIBLE-severity action hits a hard stop before it runs — no probability estimate gets to argue its way past confirmation. risk_action() collapses severity × probability into one of HARD_STOP / CONFIRM / LOG_ONLY instead of two numbers a human reconciles by eye while the moment passes. Every call — blocked or executed — appends to an audit log; 38 real events logged so far, schema: {action, predicted_severity, predicted_probability, drift_detected, status}.
Code: https://github.com/soulinpsyabstract/sipa-os-governance · Dataset mirror: https://huggingface.co/datasets/SoulInPsyAbstract/sipa-os-governance · Commits: cd442c0, 0d8a3ed · 10/10 self-tests passing.
No artifact → no claim.
repliedto sergiopaniego's post about 12 hours ago
catching up on some bookmarked reads from the summer, reading Antidoom from @liquidai
small reasoning models get stuck more easily when the task involves a long thinking trace and a hard problem. It starts repeating the same word over and over again ("Wait", "Alternatively"…), each repetition makes the next one likelier, and the generation is spent before it reaches an answer
they measured it, 10.2% of completions for an early LFM2.5-2.6B checkpoint and 22.9% for Qwen3.5-4B at greedy. After training those drop to 1.4% and 1.0%
the fix is FTPO (final token preference optimization). What I like is how narrow it is, it only touches the single token where the loop starts
three ways it differs from DPO:
> trains one token position, mid-generation, instead of whole sequences
> spreads probability across ~20 plausible alternatives instead of swapping one overtrained token for another
> keeps the regularizer in logit space, no softmax, so the rest of the vocabulary stays put
the third one is what makes it usable. If you want to edit one position without disturbing the model, you can't have a loss that reshuffles the other 150k logits on the way
and their explanation abt the result: the training teaches the model nothing new about math or code, it clears the failure mode that was blocking answers the model could already produce
full blog > https://www.liquid.ai/blog/antidoom
FTPO itself comes from Antislop, where it was built to strip overused phrasing. LiquidAI retargeted it to doom loops
and under the hood it's a subclass of TRL's DPOTrainer with compute_loss overridden, around 90 lines of loss and no new trainer
we documented that pattern in TRL's docs
https://huggingface.co/docs/trl/main/en/customization#change-the-training-objectiveOrganizations
None yet
liked a
model 14 days ago
liked a
model 27 days ago
liked a
model 7 months ago
