Spaces:
Running
Suspected Engagement Manipulation by AutoTrust AI
I suspect AutoTrust AI is artificially inflating its downloads metric here on Hugging Face.
There is a massive discrepancy between AutoTrust's download numbers and heavily backed models from prominent organizations like Cloudflare and Perplexity on Hugging Face, as indicated by these affected model searches by tag:
- https://huggingface.co/models?other=decision-model
- https://huggingface.co/models?library=safetensors
- https://huggingface.co/models?other=system-one
AutoTrust is actively touting these #1 trend rankings on X:
- https://x.com/AutoTrustAI/status/2108057938820804796
- https://x.com/AutoTrustAI/status/2107656483332915225
- https://x.com/AutoTrustAI/status/2107307463649452446
However, their X account also shows signs of inauthentic engagement -- maintaining high view counts alongside near-zero comments or organic community interactions:
Furthermore, organic Google search volume and external community traction (e.g., Reddit) for these multi-million downloaded models are virtually nonexistent.
Hi Alex,
Thanks for raising this publicly. Questions about download integrity are fair, and we’d rather answer them in the open.
1)Downloads. AutoTrust has not used bots, paid download services or any other means to inflate Hub counts. Hub downloads count requests for model files, so production and evaluation pipelines that re-pull weights add to the total. Examples include autoscaling vLLM pods, CI jobs and containers without a shared cache. Decision models are mostly called by software rather than discussed by people. That is also why download counts don’t track Reddit threads or search volume. We’d welcome Hugging Face reviewing the download logs for our repos, and we’ll post the outcome here. 2) Independent evidence. The best outside check of whether developers have a reason to use these models is a benchmark we don’t run. The Jev Decision Index is a community-maintained index of open Jev reproductions, run by @multimodalart . It combines public benchmarks with private tests of the same skills and private tasks from new domains. JEV-27B-VL led its vision board until today, when Perplexity’s Decider v1.1 moved ahead (70.6 vs. 69.6). That’s a useful reminder that the index isn’t ours, and that this category is moving fast. 3) On the comparison with Cloudflare and Perplexity. That comparison assumes the bigger budget decides who leads. On this index, the top of the vision board has just changed hands between Perplexity and us, with 1 points between the two models. Decision models don’t need frontier-lab compute. Our Blocks of Experts recipe keeps the pretrained backbone frozen and trains only a small decision block. For JEV-9B, that block has 40.2M parameters, about 0.5% of the 7.9B backbone, and was trained in roughly three B200 GPU-hours, as disclosed on its model card. Similarly, we only a fraction of JEV-27B-VL and GEV-26B-Decide backbones, and it only took less than 12 B200 GPU hours. We’ve submitted our next model to the index and will let the maintainers’ results speak when they’re posted.