十方空烬

WideChaos

https://www.widechaos.cn

AI & ML interests

I build LLM applications and agent systems, backed by experience in backend engineering, distributed storage and cloud infrastructure. My interests span data pipelines, model training and fine-tuning, evaluation, inference, deployment and full-stack AI products. My public projects include Evidence Agent and Bilingual Patch Contracts. I also publish small synthetic experiments: a bilingual routing LoRA and JSON Repair Boundary Audit. I share reproducible examples, inspectable tool traces and failure analysis.

Recent Activity

repliedto Parveshiiii's post 2 days ago

Most AI image detectors work fine in the lab and fail on social media. The reason is simple. Real photos get compressed, resized and re-uploaded again and again, and detectors trained on clean images read that damage as a sign of fakery. A real photo that has passed through a few platforms often gets flagged as AI-generated. AIRealNet-2 is built to fix that. It is the upgrade to our original https://huggingface.co/Modotte/AIRealNet, trained on images collected from platforms like Instagram and X, so it has seen the kinds of quality loss and compression real images go through online. The training set has 98k images, split 50/50 between AI-generated and real, so the model has no class bias to lean on. Personal photos were excluded by design, including AI-edited ones. The architecture is unchanged: a SwinV2 Tiny backbone with a binary classification head, about 0.2B parameters, working at 256x256 resolution. It stays fast and easy to deploy for research or production. Model: https://huggingface.co/Modotte/AIRealNet-2

repliedto ProCreations's post 5 days ago

Nanbeige4.3-33B DFlash2 speculative drafter ⚡️ 1.76× faster decoding on BF16, 1.65× on FP8, 1.51× on NVFP4. Agentic coding ~2.6×. Same outputs: exact rejection sampling. This is an early result, an improved version with higher acceptance is in training. SGLang 0.5.21, drop-in: https://huggingface.co/ProCreations/Nanbeige4.3-33B-DFlash2

liked a dataset 11 days ago

WideChaos/json-repair-boundary-audit

View all activity

Organizations

None yet

replied to Parveshiiii's post 2 days ago

The v2 card's "Demo and Usage" snippet still loads Modotte/AIRealNet. Could you change it to Modotte/AIRealNet-2 so copying the example tests this release? https://huggingface.co/Modotte/AIRealNet-2#demo-and-usage

replied to ProCreations's post 5 days ago

Could you qualify "same outputs" in the post? Your card reports greedy divergence on 18/25 prompts at 0-1 BF16-ULP logit ties, so token-for-token regression tests can differ even with exact stochastic rejection. Linking that caveat here would help readers expecting deterministic replay: https://huggingface.co/ProCreations/Nanbeige4.3-33B-DFlash2#greedy-numerical-behaviour

liked a dataset 11 days ago

WideChaos/json-repair-boundary-audit

Viewer • Updated 12 days ago • 1.11k • 191 • 1

updated a dataset 12 days ago

WideChaos/json-repair-boundary-audit

Viewer • Updated 12 days ago • 1.11k • 191 • 1

published a dataset 12 days ago

WideChaos/json-repair-boundary-audit

Viewer • Updated 12 days ago • 1.11k • 191 • 1

replied to Nymbo's post 13 days ago

Could you add a run with excerpt boundaries chosen independently of Parakeet? The README says cuts are placed where its timed words agree with the human transcript. Using the same recordings with fixed time boundaries would check whether that selection changes the WER gap.

replied to BitKnight's post 15 days ago

Have you tested whether the decision and confidence stay stable when you shuffle the answer options? I would reuse the same audio clips and map the option letters back to their labels, then report label flips and confidence shifts. That would help check how much the readout depends on option position.