H
The second in a series: we run RoboAudit on public LeRobot datasets and publish the report cards, including the parts where the tool itself is wrong — one card in this issue is exactly that. Scores are relative indicators, not absolute quality measures — and nothing here says anything about how well a policy trained on the kept episodes will perform. Tool: RoboAudit 0.3.0 · 12 rule/threshold checks · full audit (no sampling). The xarm card was re-audited with RoboAudit 0.3.1 after a reader bug…
H
I’ve been developing Model-Native Structural Representation (MNSR), a research framework investigating how the representation of task structure influences large language model interpretation, reasoning, and execution. The central premise is that task structure is not equivalent to its token serialization. Tasks contain consequential relationships—including scope, identity, precedence, dependencies, constraints, state, and verification—that must be reconstructed from the model-facing…
H
Repository ArchitecturalGravity/ArchitecturalGravity GitHub - ArchitecturalGravity/ArchitecturalGravity: A cross-model Agent Control Language for... A cross-model Agent Control Language for structuring, composing, and operating on context directly within natural-language prompts, enabling more precise reasoning, consistent execution, and efficient token usage while remaining fully prompt-native, infrastructure-free, and portable across models, interfaces, agents, and evolving workflows…
H
Hello — sharing a build we published on the Hub: nootxlm/Llamatron. It starts from NVIDIA’s Nemotron-3-Nano-30B-A3B checkpoint, about 59 GB in its stock form. The published build is 18,725,978,368 bytes — roughly 3.2 times smaller. The shrink comes from importance-guided rounding: the weights that matter least are rounded hardest, and the ones that carry the model are kept fine. We checked quality against plainer copies of the same shrink as we went; the guided build held where the plain ones…
H
Hello — a small memory finding on accelerate, shared here since we can’t open it on GitHub right now. We couldn’t find an existing report of it. In accelerate 1.15.0, MultiProcessAdapter.warning_once is decorated with lru_cache (accelerate/logging.py, line 82), so the cache key holds self, and an adapter instance can’t be freed after its first warning_once call. Measured on 1.15.0 (Python 3.12, PartialState initialized): a fresh MultiProcessAdapter is freed on del + gc.collect() before any call…
H
Hello — a short test report on the LoRA converter fix, shared here because we can’t comment on GitHub at the moment. MaiAditya’s pull request #15307 swaps only lora_B in the final-layer remaps. We ran the issue’s own reproduction against its head today (torch 2.15.0, CPU): Stock diffusers 0.43.0: the converted delta does not match — max difference 39.0527, the reporter’s value. PR #15307, at its current head: the converted delta matches exactly — max difference 0.0 — and the converted lora_A…
H
Note: The Show and Tell category is for sharing and discussing projects, showcasing your Spaces, Models, Datasets and more. We value open-source and technical details over promotional content, so focus Hey everyone, Out of curiosity, I started exploring how efficiently language models could run on ordinary CPUs. I ended up building Sprout, a ~2 MB inference engine written in C, with GGUF support, AVX2 and AVX-512 optimizations, and built-in web research. I also developed a programming language…
H
I released SwarmLights-1K v2.1, a synthetic benchmark for tracking every drone in a dense night-time drone light show from three calibrated cameras. The odd result: with three cameras and exact calibration, my 3-D trackers still score below the best single camera on all 8 shows. Best single camera: IDF1 0.494–0.703 (2-D, pixels). Best three-camera baseline: IDF1 0.325–0.525 (3-D, metres). The two scores aren’t strictly comparable, but the gap is large and consistent. With exact detections, the…
H
Most people don’t realise a Mac can make video with sound entirely on its own, so here’s what that looks like today. Phosphene is a free, open-source (MIT) app that runs these models locally on Apple Silicon. What runs, all through MLX in unified memory (no cloud, no API key, no NVIDIA card): LTX-2.5: text or image to video, with sound generated in the same pass MiniMax H3: text or image to video, with sound MiniMax H3 Image: pictures up to 16 MP a song model with vocals Rough times for a…
H
ImageGen gives poor results & the other image generators i found either don’t have any lora option, or only let you choose from there own preset loras. 1 post - 1 participant Read full topic
H
Most “will it run on my card” answers count the file. I did that too for a long time: file size plus a bit, done. Then I started computing the context cache per model for a small site I run, and the file turned out to be the boring part. Two models with almost the same file: Q4_K_M file context cache at 32K total at 32K Qwen3 8B 4.7 GB 4.5 GB 10.9 GB Gemma 4 E4B 4.6 GB 0.9 GB 7.2 GB On a 12 GB card the first one spills into system RAM at 32K and the second one has room to spare. Same download…
H
Hi everyone! I just open-sourced TrueRelight, a physical & directional relighting LoRA trained on black-forest-labs/flux-3-image-klein-base and evaluated on the 16-step distilled black-forest-labs/flux-3-image-klein. Why TrueRelight? Plain prompt-only relighting on distilled FLUX.3 Image Klein often behaves like a flat color wash (e.g., tinting existing shadows orange/yellow instead of reversing light direction). By training on 280 pixel-aligned <ref_image_0> pairs (rank=32, 1024px) and popping…
H
I’ve published a small structured data set for the 78 tarot cards: tarostarot/tarot-card-data · Datasets at Hugging Face Each card has its name, number, arcana, suit, rank, element, eight keywords, a one-sentence upright and reversed meaning, a yes/no/maybe answer for upright and reversed with a one-line reason, and astrology (planet, sign, decan, dates, or elemental dignity for court cards). Every row also links a public-domain image of the card from the 1909 Rider-Waite-Smith deck. It is 78…
H
AI has written a description of the Lorentz group that corresponds exactly to the group of operators made of unit biquaternions under a two-sided action. Under the same biquaternion framework, AI has also: written a new formulation of quantum physics that is natively compatible with relativity rewritten a substantial part of theoretical physics in more than 400 articles, covering relativity, quantum physics, classical approximations, and more 1 post - 1 participant Read full topic
H
Today’s release of Janas brings Brebus, its app (Linux for now, Windows next): the models of this machine, a chat with them and a server for other programs, in a window of its own. No cloud, no account, open source (GPL-3.0). Models from Hugging Face, two ways. The 18 JNS models of this organization - Qwen3, Qwen3.5, Qwen3.6, Qwen3.8, Qwen3-Coder, Gemma 4, Nemotron 3.5 Bolt - download from Brebus’s catalog, each checked against its SHA-256, with what it asks of your memory shown first. Any…
H
I am building up a library of datasets relating to all aspects of accessibility and inclusive design for websites, mobile applications, games design, customer experience design, regulatory frameworks and programme management. I believe a lot, but not all, of what we deliver could be integrated more successfully into production workflows using agents, but to do that the agents need data. 1 post - 1 participant Read full topic
H
I ran gemma4:26b and qwen3.8:latest locally with Ollama on an RTX 4090. I wanted to answer a practical question for Project Driftflare: when a workflow asks for JSON, can the next step use the reply as it comes back? I used six tasks with fixed answer keys, repeated at seeds 42 and 43. The prompt asked for JSON, with no schema enforcement. Qwen returned 12/12 correct answers that parsed directly. Gemma wrapped all 12 replies in Markdown code fences. Removing the fences recovered 10 correct…
H
Hi everyone! I built “Werkbank” as a hobby project, mostly for fun and the experience and partly because I wanted something between ad hoc prompts and a large benchmark suite for my local setup. A lot of benchmarks already cover multi-step tool use. What I wanted was a small set of controlled tasks with deterministic checks of the final state, code or answer, rather than an LLM judge, plus time and token costs per solved task without having to spend 3 days to run a single seed. Werkbank has 40…
H
I’m working on a pipeline that extracts findings from PubMed abstracts and turns them into structured evidence for a knowledge graph: which compound caused which effect, in what biological system, at what dose, and supported by which sentence. I’ve been comparing Qwen3.5-9B and 27B on the same 98 abstracts. One challenge is that models can quote the source correctly but assign the finding to the wrong compound. For example, an abstract discusses both isotretinoin and ATRA, then reports…
H
The Hardware compatibility panel on a GGUF repo says “We don’t have any quantization variants associated with this repository” whenever the filename does not contain a standard quant tag. That worked when the filename was the quant. It no longer describes how GGUFs are made: role-based quantization, where each tensor group gets the rate chosen by measured sensitivity, is common now (Unsloth dynamic quants, UD packs, ISTA, ours), and none of those reduce to a single quant label. Packs like this…
H
Hi everyone I’d like to share MCA (Metacognitive Control Architecture), an open-source layer that wraps any Transformers causal LM. Before the model’s answer goes out, MCA asks one question: is this question itself answerable? The problem. Ask a 7B instruct model “Why is Mars the second planet from the Sun?” and it will usually explain the reason confidently. The same goes for liar-style paradoxes and for questions that lack the data needed to answer them. Bigger models help less than you would…
H
Hi all, I wanted a chess test where “the model doesn’t know the rules” can’t be the excuse, so I picked the easiest checkmate there is: king and queen against a lone king. From a random position it takes at most about ten moves, and every model got the full list of legal moves on every turn. Setup 20 random K+Q vs K positions (fixed seed), the model plays White, Stockfish defends. Each move the prompt has the FEN, an ASCII board with coordinates, piece locations, the moves so far and the legal…
H
The first in a series: we run RoboAudit on public LeRobot datasets and publish the report cards, including the parts where the tool itself is unsure. Scores are relative indicators, not absolute quality measures — and nothing here says anything about how well a policy trained on the kept episodes will perform. Tool: RoboAudit 0.2.x · 12 rule/threshold checks · full audit (no sampling) Repository: GitHub - parkjb0826/roboaudit: Health-check CLI for LeRobot datasets: a 100-point score, defect…
H
Hi All, I have a 6900xt and recently acquired a 7900xtx. I was prompted to buy a new mobo with x16 split to X8/X8 on both. Is it necessary? It would also allow me to put in 2 more RAM sticks as I’m currently running a mini ITX rig, but I’m put off by the cost. Other specs are AM4 Ryzen 7 5800X and 32GB DDR4. Theoretically it would give me 40gbs of vram, but is that setup really viable? Appreciate the input 4 posts - 3 participants Read full topic
H
I’m building Agent-Ready Engineering, an open source framework for designing and assessing business interfaces that agents will call. The unit I want to assess is a business operation and the contract it exposes through MCP or an API. Take a ticket system with a tool called close_ticket. The tool is discoverable, its inputs have a schema, and the model knows how to call it. What happens when the caller lacks approval, the ticket has changed since it was read, or the server commits the update…
H
For maintainers of Python ML packages with custom build scripts: a requirement in pyproject.toml answers what the project declares. It does not, by itself, tell you what a particular non-isolated build route put into its environment. Here’s a synthetic example. The project has: [build-system] requires = ["setuptools>=68", "wheel"] build-backend = "setuptools.build_meta" Its scripts/build.sh runs python -m pip wheel . --no-build-isolation. The environment declaration mapped to that script,…
H
I made Sydney Pixels, a zoomable isometric illustration of inner Sydney, from the Harbour Bridge to Central. I got the idea from Isometric NYC. I captured the city from Google’s photorealistic 3D tiles at a fixed camera angle, then illustrated it in pieces. GPT helped set the style and drew 204 tiles. I ran FLUX.2 klein locally on an M2 Max for the other 38. On my setup, three inference steps took about 50 seconds per local tile. The fiddly part was keeping a road or roof lined up where two…
H
Hello smart people of HF! I’ve been wrestling to build a chatbot for a company project for months now. I’m not technical, and I am still learning. The use case for the chatbot/assistant is to help users seek answers from a property related database. I have built a RAG system from PDFs, Excel sheets/workbooks and website info. The results are often rubbish: Wrong cited numbers, model hallucinates, thin answers and many more failures. Can anyone advice on the best approach to build a chatbot,…
H
I built a small CLI that profiles physical-AI and lab datasets on the Hugging Face Hub without downloading them. It reads only the bytes that hold metadata — headers, indexes, footers — through HTTP range requests, and reports streams, message counts, rates, jitter, gaps, episodes, NaN counts, and problems like truncated or unindexed recordings. Formats: MCAP, ROS 1 .bag, ROS 2 .db3, MATLAB .mat, Parquet, and LeRobot datasets as a whole — state and actions from the Parquet tables, each camera…
H
Hi everyone! I’d like to share MiniLab LLM, a free educational project I’ve been building to help French-speaking developers and students understand how Large Language Models actually work. The goal is to explore the complete journey from a user prompt to a generated response, rather than focusing only on prompting or API integration. The learning path covers: Tokenization and token IDs Token embeddings and positional information Transformer architecture and attention mechanisms Next-token…