Hugging Face's logo
R's picture

R PRO

Twu31
10
jackhooke's profile picture ErenAta00's profile picture mihailgribov's profile picture
·

AI & ML interests

None yet

Recent Activity

repliedto mihailgribov's post about 1 hour ago
TypeCastLM: Jev-class decision models on frozen LLMs TypeCastLM is a decision model built around a frozen LLM. Given a text and a question, it returns calibrated probabilities over the allowed answers in one forward pass. The LLM is not fine-tuned: not one of its weights changes. TypeCastLM swaps its LM head for a mini one: a linear matrix whose rows are mostly the model's own output rows, with the three verdict rows (true, false, unsure) fitted. A fitted row catches an answer the model spreads over many tokens, such as "yes", "true" or "correct", where a single vocabulary row sees one word. A head that small leaves little room to overfit. It speaks the Jev API, so a Jev client switches by changing the base URL. The first model, [typecastlm-qwen3.5-3.8b](https://huggingface.co/mihailgribov/typecastlm-qwen3.5-3.8b) on Qwen3.5-4B, is among the top open frozen-4B models on JevBench v1.6. Besides Jev's yes/no, choice and scale modes it has `tfu`: yes/no with a third answer, `unsure`, for when the text lacks what the decision needs. Specs: • Size: 3.76B params; 7.5 GB bf16, 4.0 GB Q8_0 • VRAM: 12 GB is enough, 16 GB comfortable • Latency p50 (batch 1, bf16, RTX 5060 Ti): ≤200 tok 48 ms · 1k 93 ms · 4k 566 ms • Calibration (ECE): BoolQ 0.052 · RTE 0.012 · FEVER 0.019 • Context: 32k tokens • Runtime: transformers or llama.cpp, fully offline • JevBench v1.6: 19.7 `pip install typecastlm`
repliedto ErenAta00's post about 1 hour ago
Maverick 1.1 is out. When we released the first version, the feedback was clear: real people do not give one clean command at a time. They change their minds, they point and say "put that on the shelf", they ask for a lamp that is in another room, and sometimes the one speaking is a remote expert helping someone on site. 1.1 is built around that. - Say "no, the other one" and it undoes the last action and repeats it on the other object. If the effect can't be undone, like a button that was already pressed, it tells you. - It sees what people are looking at or pointing at, and who is speaking. - It finds objects in other rooms, goes there and acts, or asks which one if there are two. - It follows a short plan in one sentence: "open the cabinet, put the book and the alarm clock in it, then close it". In the clip, a remote expert corrects the model, then asks it to highlight what the technician is looking at. On test sets we sealed before training and ran once, the new abilities score 93 to 100%, where 1.0 scores 33 to 87%. On sentences written by people (the ALFRED set) the two versions are level, at 83.2% and 83.8%. Where it still falls short: in unfamiliar situations it tends to act when it should ask, and a remote speaker can act in the scene with no permission check, so your app has to enforce that. Same 4B model and licence (Apache-2.0), the same 11 tools plus `undo`, still offline on a 4 GB laptop GPU. The Unity package for the new scene format (1.1.0) is on GitHub: https://github.com/mcbu-xrlab/Maverick-Unity Model: https://huggingface.co/ErenAta00/Maverick-4B-Unity-XR-Agent-1.1 Article and full demo: https://huggingface.co/blog/ErenAta00/maverick-1-1-undo-pointing-other-rooms Built at the Extended Reality Laboratory (XRLab), Manisa Celal Bayar University: https://huggingface.co/ExtendedRealityLabMCBU
reacted to IsValorum's post with 👍 about 1 hour ago
Navigating the APEX-I and VAL-APEX Quantization Ecosystem on Hugging Face As the catalog approaches 400,000 total downloads across 69 releases, all models are now structured into dedicated official collections on Hugging Face. Our goal remains consistent: eliminate the compromise between memory overflow and cognitive collapse on consumer hardware (especially 16 GB VRAM GPUs like the RTX 4080, RTX 4070 Ti Super, and RX 7800 XT). Instead of uniform block quantization, all APEX builds use per-tensor mixed precision guided by 4-moment Importance Matrix statistics: - MiniPlus Tier (approx. 14 to 14.8 GiB): Strict 0% 2-bit policy. All MoE routed experts locked at minimum IQ3_XXS with Q4_K/Q6_K attention. Designed for 24 GB setups or 16 GB setups with moderate context. - NanoPlus Tier (approx. 11.5 to 12.3 GiB): Strategic use of IQ2_S/IQ2_XXS exclusively in core feedforward blocks, keeping routers and attention high-precision. Leaves ample VRAM for 16k-32k context on 16 GB GPUs. Explore the active collections: - MoE Standard (MiniPlus & NanoPlus): https://huggingface.co/collections/IsValorum/apex-i-miniplus-v21-standard-moe-6acfee77961c8c0cf5a083b5 - MoE Abliterated (MiniPlus & NanoPlus): https://huggingface.co/collections/IsValorum/apex-i-miniplus-v21-abliterated-moe-6acfee762fc4a8cf2298c55b - 27B Dense Hybrids: https://huggingface.co/collections/IsValorum/val-apex-miniplus-standard-27b-6acfec54ca4ce4cf5cd5bdb6 - Apple Silicon (MLX): https://huggingface.co/collections/IsValorum/mlx-apex-abliterated-6abedb1f9d400e984efe4fc6 Full technical article and sizing breakdown: https://huggingface.co/blog/IsValorum
View all activity

Organizations

None yet