XIO

xWire

Posts from every feed QUEST has found across the web — blogs, news, podcasts, channels and more.

Hugging Face Forums - Latest posts

Latest posts

Phys-ai-dataprofiler — know whether a dataset is good enough before you download 500 GB

The overlap check is the right call, and your correction stands: a uniform shift leaves every nearest-sample gap unchanged, so record time alone can prove range overlap, never offset. For the header check, the series to report is header stamp minus record time for each message — a constant offset shows as a constant difference between two streams’ series, drift as one that widens steadily, and both stay visible even where the ranges overlap fully. A fixture with one stream’s headers shifted by…

Detect and display GGUF quantization from tensor info, not from filenames, especially helpful for role aware

@BoxyFit The current reliance on filenames fails because heterogeneous quantization assigns distinct precisions to specific tensor groups based on sensitivity, not a single global label. When parsers expect a uniform quantization flag, they miss the per-tensor `ggml_type` stored in the header. A robust implementation must iterate the `tensor_info` array, aggregate the histogram of types, and calculate the weighted average bits per parameter. This approach correctly handles mixed-precision packs…

Phosphene: LTX-2.5 and MiniMax H3 video, running locally with MLX on a Mac

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…

Vetting an open model before deployment

How would you turn this into a practical checklist for a founder deciding whether to use an open model? Beyond testing its answers on real tasks, what should they check about the source of the model and any code it asks you to run, the license, and where user data goes? And how do you decide which uses are safe enough to try, which still need a person checking the output, and when not to use the model at all? I’d be interested in a simple example rather than just a benchmark score.

MCA: open-source neuro-symbolic wrapper that teaches 7B LLMs to challenge false premises, explain paradoxes and abstain (code + dataset + Space + 2×T4 benchmark)

That’s a cleaner design than growing the held-out set, and the DeepSeek numbers fit it exactly: those four traps were caught at the default and all four slipped through at 0.9. One thing I’d add: tuned on rebuilt traps alone, τ would just slide towards “abstain on everything”. So I’d tune it jointly on the rebuilt traps and on control facts that also reach the entropy stage. Almost all controls already get there, since R1–R7 rarely fire on them. The objective becomes: the highest τ that still…

MCA: open-source neuro-symbolic wrapper that teaches 7B LLMs to challenge false premises, explain paradoxes and abstain (code + dataset + Space + 2×T4 benchmark)

Only the traps the earlier rules let through can calibrate τ — on held-out DeepSeek that set was four traps, all caught at the default and all missed at the tuned 0.9, so the 100-item benchmark was never really voting on τ at all. Growing that set is cheaper than growing the held-out set: take traps the symbolic layers already catch, change only the feature the catching rule keyed on — the flagged referent, the premise’s surface wording — while keeping the question itself unanswerable, until…

A compass gate for self-updating models: learn by day, consolidate by night

Rick, you asked whether “a judge that understands meaning” does better on the pairs where the answer should flip. On these examples, a language-model judge given the rule in words did, though the comparison isn’t like for like (see below). We wrote the plan down and committed it before anything ran. The judge. Cael, a 27-billion-parameter language model running on our own machine. When asked, Cael agreed to take part. Tetsu, our smaller model, said no, so isn’t in it. Each of the 26…

MCA: open-source neuro-symbolic wrapper that teaches 7B LLMs to challenge false premises, explain paradoxes and abstain (code + dataset + Space + 2×T4 benchmark)

@nootxlm, as promised: the uncensored audit run is done. It used Kaggle 2×T4, 3 models and audit_all_layers=True, so every layer ran on every item. Entropy is now observed on 100/100 benchmark items and 24/24 held-out items per model, compared with 20–26/100 before. Per-rule verdicts on the 100 items (each rule evaluated on its own, not first-match; “ctrl” = control facts where the rule fires): Rule Qwen2.5-7B DeepSeek-R1-7B Llama-3-8B R5 NLI contradiction 13 (ctrl 0) 6 (ctrl 0) 16 (ctrl 6) R7…

Same file size, five times the context memory

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…

How do you design memory systems for long-running AI agents?

To keep it from becoming taste one level up, here is the profile in numbers. In the ablation behind our first warning, moderate load meant decay 0.15 per day, a 45-day horizon, 300 facts and 100 queries; the harsh profile raised these to 0.25 per day, 60 days, 500 facts and 150 queries, with the components unchanged (abstract of version 9, linked above). One correction to our earlier post: under that profile the shifts were not tiny; five components fell from exactly 0% to below -89%.…

Dual AMD GPU setup

Can’t say much about the board, but on the “is 40 GB worth it” part I have numbers. I keep a table of what open models need in memory (Q4 GGUF file + context cache + a bit of runtime overhead), so I ran your two cases through it. The 7900 XTX on its own, counting about 21 GB as really usable: 27 of the 47 models in my table fit at 8K context and 24 at 32K. The biggest ones that still go on the single card are things like Gemma 4 31B or Qwen3-Coder 30B, and those are already tight. Both cards,…

AI Model Radar – measured hardware fit for open models (real GGUF sizes + per-architecture KV cache), with a JSON API and MCP server

Yes to both. Good question btw, nobody had asked about the schema side yet. The MCP tool (find_fit) just hands back what the REST endpoint returns, so the three parts are separate fields, all in GB, and there’s a basis block that says which one is measured and which one is only assumed. Qwen3 14B on a 4070 at 32K: "label": "OFFLOAD_REQUIRED", "quantization": "Q4_K_M", "estimateGb": { "weights": 8.4, "contextCache": 5, "overhead": 1.9, "total": 15.3, "usableMemory": 10.6, "offload": 4.7 },…

Gemma got 10/12 answers right, but none of its raw replies parsed as JSON

Good catch. I called it an ordering miss, but that lumped two things together. At both seeds, Gemma returned "Design, Build, Test". If I split that string on commas and trim the spaces, the sequence scores 2/2. If the next step expects answer to be an array of strings, the same replies score 0/2 on type. With the schema, Gemma returned ["Design", "Build", "Test"] at both seeds: 2/2 on sequence and 2/2 on type, and both raw replies parsed directly. The schema left the ordering open, so it…

TrueRelight: Directional & Physical Relighting LoRA for FLUX.3 Image Klein (16-step distilled, Open Weights + 280-Pair Dataset + Interactive Space)

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…

Tarot card data for all 78 cards (CSV/JSON, CC BY 4.0)

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…

Benchmarking System One small LLM decision APIs vs. in-process tree ensembles and exact TreeSHAP on 64-feature tabular payloads

Reproducibility is the part I would underline too, and it is stricter than people expect. With a fixed tree structure the SHAP values are exact and deterministic, so a decision record can store the model artifact hash, the payload, and the attribution vector and re-derive the same numbers years later. A sampling endpoint cannot promise that, and a reason code that rewords itself is the first thing an auditor asks about. Putting the LLM after the decision, as prose over fixed attributions, keeps…

Overcoming False Positives in Real-Time Computer Vision for Engine Room Oil Leak Detection

For now, I looked into some related research and ran a few small checks with synthetic images: Short answers to your three questions Your current combination of static ROIs, pixel-change/background subtraction, semantic segmentation, and person detection with ByteTrack is a useful starting point. I would keep it as a candidate generator and separate what the system subsequently needs to establish: (1) whether an image observation is geometrically and photometrically trustworthy; (2) whether a…

Questions about GGUF

Hmm… It might be easier to think of GGUF as a file format, the software that uses it, and quantization itself as three separate things: The short answers 1. What are the downsides of GGUF quantization? The usual tradeoff is not caused by the .gguf container. It comes from reducing the precision of the model’s weights. A Q4 or Q8 variant generally stores approximations of some weight values, making the model smaller and often easier to run locally. Those changes can also affect next-token…

I made a web sdk with chrome extension to airgap open weights and with open web use

Words cannot describe how happy I am this worked. I need to refine the UX but the core works. I can go site to site online offline and AI works to connect to web apps with NO API COST or middleman… Free Artificial Intelligence for Chromium When I get finished testing it will be open source. If you think about what I made it means you can now use local AI with app interaction with no egress.

KoFinClaim-Abstain: a Korean benchmark that scores LLMs by the cost of false accepts on financial numbers

Thank you for the careful replay — independently re-running score.py on all seven baseline files and checking the 500-row hash is exactly the kind of check I hoped someone would do. On the two rows, I re-read the claim and evidence and I agree with both readings: s4-T025: the claim is a negative statement (“EPS 700 cannot be established from this data”), and the evidence explicitly lacks the weighted-average share count. Under the dataset’s own convention (s5-V019, s3-P014, s4-T010) this is…

The AI safety question may be backwards

Agreed, One of the design problems I run in to a lot in my interactions with the large models is a presumptive authority to act. When they are not certain they are still designed to give the answer. I enquire about the feasibility of a script and it produces the script without asking if that is what I need. – My current working solution is in attempting to encourage ‘alignment’ prompts through the conversations and rules around asking questions when dealing with uncertainty. - If they included…

Questions about GGUF

This is kinda an exploratory question. I dont really know a lot about GGUF and the current agent space. I just started messing with some agents for fun and math purposes. What are the downsides to GGUF quantization? Would it be better to have exact compression? Like GGUF without the quantization? Just wondering. Maybe I dont understand what GGUF does. Or is there some inference benefits there?