H
I have gone through the examples on how to extract data from Tim-Pinecone/LOC-Chronicling-America dataset : README.md · Tim-Pinecone/LOC-Chronicling-America at main Are we getting the individual articles or the entire news paper? Inspecting some outputs I don’t see any “boundary” of articles. So, I am not sure that the outputs are at the article level. If not then How to narrow down the search to get articles on a topic - e.g., earthquake, suffrage etc? Has anyone done this? Thanks!! 2 posts -…
H
I’m building Spore, an Apache-2.0 runtime for persistent AI agents. Here’s an experiment using it to build and test a small Godot game. The agent wrote the game, generated its art and sound in code, and wrote bots to test its balance over seeded runs. It passed the first prompt’s targets—but reviewing the results exposed gaps those targets didn’t cover. A second prompt added requirements. The agent improved its search, revised the bargain mechanic and fixed the ending. Nobody edited the game by…
H
I’ve spent about a year building Neuro, a programming language for AI work. You write the model, the training loop and the GPU code in one language, and the compiler produces a single native program. Whoever runs it doesn’t need Python installed. This small program trains on the CPU, then predicts on the GPU: @grad func loss(w: &mut Tensor<f32, [2, 1]>) -> Tensor<f32, []> { val x: Tensor<f32, [4, 2]> = [[1.0, 2.0], [2.0, 0.0], [0.0, 1.0], [3.0, 1.0]] val y: Tensor<f32, [4, 1]> = [[5.0], [2.0],…
H
The certificate generation website has error .. Any fixed schedule? 2 posts - 2 participants Read full topic
H
Hi Hugging Face team, I’m an author and the submitter of Open-HumanWAM: https://huggingface.co/papers/2611.06817 We intended to submit the updated v2 version, but accidentally submitted while the paper was still showing v1. Could you please withdraw the current Daily Papers submission so we can resubmit once v2 is available? We would like to keep the underlying paper page and associated repositories intact. If the existing submission can be updated to v2 without withdrawing it, please let us…
H
Hello Hugging Face team, I am following up on a moderation review request sent to safety@huggingface.co on November 5, 2026, concerning a private Docker Space under my account, @robthepirate. The Space URL and repository revision are included in that email. The build completed and the container logged application startup at 05:26 UTC that day. At 05:35 UTC, the runtime API reported PAUSED with “Flagged as abusive.” The response available to me did not identify a specific rule or component. The…
H
I’m interested in how far we can leapfrog computation by treating computation as cache—especially when the process that discovers the next compression becomes something we can compress. In his Stanford HAI conversation, Jeff Dean questioned whether a recursive improvement loop could run infinitely fast, given the physical work required. My question concerns how recursively learned compression changes the work required by subsequent loops, including the work of producing the next compression. To…
H
When I’m iterating on a small model or RAG setup, the hard part isn’t the training loop — it’s remembering why I chose a particular dataset split, which prompt variants I already tried, and what the eval notes meant two weeks later. Right now I bounce between notebooks, a scratch doc, and chat logs, and the “why” of a run gets lost. How do you keep experiment notes, data decisions, and eval commentary attached to the actual work so a future-you (or a teammate) can reopen it without archaeology?…
H
A model that works in a notebook demo isn’t necessarily ready for production. Before deploying a model from the Hub, start with the model card: it describes intended use, known limitations, license terms, and sometimes training data notes. Then test it on a sample of your own data rather than relying on leaderboard scores. Check how it handles long inputs, other languages, and out-of-scope requests, and measure latency and memory use under realistic load. Keep the previous version deployable so…
H
An eval is a repeatable test for an AI system. You collect example inputs, decide what a good answer looks like, and score the outputs the same way every time. That lets you compare two prompts or two models with numbers instead of impressions. Scoring is usually a mix of programmatic checks (valid JSON, required fields), human review, and an LLM acting as judge. LLM judges are fast but tend to favor longer or more confident answers, so it helps to check them against human ratings. Hugging…
H
Red teaming is deliberately trying to break your own AI system before someone else does. For language models, that usually means testing prompt injection, jailbreaks, attempts to extract the system prompt, and requests that push the model toward harmful or off-topic answers. It works best when it’s repeated. Models, prompts, and retrieval data change, so an attack that failed last month can succeed after an update. Most teams keep a list of attacks that worked and re-run them after every…
H
SelMem: From Reconstructive Memory to Behavioral Learning A story from the first 100 commits of my project, SelMem, covering September 13 to November 2, 2026, supplemented by experimental reports and associated code changes. The Common Thread The story of SelMem is one of a hypothesis progressively confronted with experimental reality. Initially, the intuition was clear: If two identical LLMs experience different events, then selectively remember, forget, and reconstruct those experiences, they…
H
huggingface.co LuzCommand/lil-beastie-4b · Hugging Face Text Generation model with 4B parameters. Download, explore files and run it on Hugging Face. Here is my first go at a small terse abstainer. Just curious if anyone else is trying to do this? 8 posts - 5 participants Read full topic
H
I’m the author, so read this with that in mind. ScopeHaven runs an agent’s tasks against a local copy of an app, as different users, and checks whether “done” matches what the app’s own records show. No model takes part in a verdict, and when it can’t tell, the row says inconclusive instead of pass. The quickest demo needs Node 22 or 24 and nothing else: git clone https://github.com/YusefSyed/scopehaven-oss.git && cd scopehaven-oss ./scopehaven assess --target demo-vulnerable That’s a synthetic…
H
I’ve published six small text tools for writers. They run entirely in the browser: no model, no server, nothing stored. Two scan automatically: Scene Checker flags words that name an emotion directly and similes (dictionary-based, English and Turkish). Text Analyzer shows sentence length distribution, dialogue share, repeated phrases and moving-window lexical diversity. The other four don’t score anything: the writer labels (narrative distance, story opening) or enters details (subtext, planted…
H
I recently tried Agaram 2.5B Instruct from Murai Labs locally through Ollama. A Tamil-first model caught my attention. I wanted to see how it handles everyday expressions—particularly the differences between Sri Lankan and Tamil Nadu Tamil. The experiment I put together 20 everyday phrases, each with a Sri Lankan Tamil and a Tamil Nadu Tamil version expressing the same meaning. I asked Agaram to translate all 40 sentences into English. I used the Q8_0 GGUF with temperature 0, graded the outputs…
H
Hello Hugging Face community! We are introducing Ryzova Alphi v1, a publicly available AI model release from Ryzova, developed using Google Gemma 4 E2B-it as its foundation. Alphi is part of Ryzova’s broader research into local-first AI, structured memory, adaptive intelligence, and agent-based systems. Model release The Hugging Face repository includes: FP16 Safetensors model weights F16 GGUF for full-precision local inference Q4_K_M, Q6_K, and Q8_0 GGUF variants for different hardware…
H
I’m building an OCR pipeline for invoices, delivery notes and receipts (native PDFs, scans, photos) that exports one Excel sheet per page: header fields, the item table as a real table, and totals. Every value carries a status: reliable, to verify or absent. The goal is that it never presents a guess as a certain value. Two things I’d like your opinion on: How do you evaluate table reconstruction on scanned registers (40+ rows, 4 columns)? Mine sometimes keeps only a partial block, so I added a…
H
Hello Hugging Face community! I’m Devanandha VS, an independent AI researcher and MSc Artificial Intelligence graduate from Ulster University, UK. I’m seeking independent researchers, computer-vision practitioners and open-source developers interested in evaluating an agricultural AI research project. Project Overview Wheat Rust AI Detection is an open-source deep-learning project using a MobileNetV2-based classifier to identify three wheat leaf categories: Healthy Brown Rust Yellow Rust The…
H
Follow-up to last month’s post. Every item below is a row in our open price-change log, checked against the provider’s official pricing or deprecation page. Price increases already announced (not applied yet) Gemini 4 Argon: input $2 → $4, output $10 → $20 per 1M tokens from Dec 1 Gemini 3.6 Flash and Gemini 3.8 Flash: input $0.75 → $1.50, output $3.75 → $7.50 from Jan 1, 2027 If you budget for next quarter on Gemini, use the new numbers. Worked examples of what the increase does to a monthly…
H
Hey everyone, I want to add a real time chat and video to my app. I’ve been checking a few providers, and the most popular ones are Getstream, Sendbird, MirrorFly, Agora and Ably. Before I invest any of them, I’d like to hear from people who’ve used them. Which one would you recommend and how was your experience. 6 posts - 5 participants Read full topic
H
What can silently go wrong between an SFT JSONL file and the loss tensor? When an LLM fine-tuning job crashes with a CUDA OOM or produces NaNs, the failure is obvious. The more insidious problems happen when a run completes cleanly with exit code 0, the loss curve descends predictably, but the loss tensor is not optimizing what you intended. Between raw training records on disk and the computed loss tensor on GPU, a training framework executes a multi-stage transformation: Raw records →…
H
I’ve been working on the Web Models API, a proposal that lets websites request specific open-weight models and run them on the user’s device. The browser manages the model download, storage and execution, while users control which sites can access each model. The goal is to give developers a way to use the models they’ve built and tested against without managing the inference runtime in their application. Model resources can be shared across sites, and prompts don’t need to be sent to an AI…
H
Hi Hugging Face team, I’m an author of arXiv:2610.35497. The paper went live late in the November 4 Daily Papers feed and received little visibility. Could you please withdraw the Daily Papers submission so I can resubmit it on a fresh day? Please keep the existing paper page and its GitHub and project website links. Paper page: Paper page - When Fine-Grained Credit Assignment Helps and Hurts: Multimodal Agentic Reinforcement Learning through Video Editing Thanks for your help! @AdinaY @hysts…
H
Hi all — I measured how many tokens Urdu costs compared with English for the same meaning, and wanted to share the numbers and ask for feedback. I built a small parallel corpus (30 sentence pairs, 17 domains, CC-BY-4.0) and tokenized both sides with three tokenizers: Tokenizer Vocab English tokens Urdu tokens Inflation GPT-2 ~50k 320 1,999 5.39× Mistral v0.3 ~32k 359 1,554 3.73× Qwen2.5 ~151k 318 1,002 2.72× Per character it’s starker: English gets ~5.35 characters per token on GPT-2, Urdu gets…
H
Hey everyone! :3 I’ve been working on a project called Hierarchos-Native for quite a while now, and I figured I’d finally share it here since a pretty big part of what I’m doing revolves around Hugging Face models and infrastructure! It’s an open-source machine learning backend I’ve been building in Rust, using Vulkan compute for transformer training, fine-tuning, and inference. The whole thing started because I’ve always been a huge nerd for low-level programming and squeezing performance out…
H
A compass gate for self-updating models: learn by day, consolidate by night Hi all. I’m not an engineer, but I’ve been thinking about self-updating models like MIT’s SEAL and wanted to share an idea for feedback. The idea: a model gathers experience by day while its weights stay frozen. Each night it drafts an update from that experience, rehearsing old skills so it doesn’t forget them. Before the update goes live, it has to pass one fixed principle, “treat every other as yourself,” plus a…
H
Suppose a support agent may close a ticket only when the customer confirms resolution. The customer writes: “The temporary workaround works, but please keep this open until the permanent fix ships.” A classifier emits customer_confirmed_resolution = true. A deterministic policy gate checks the label and permits closure. The audit log says the policy check passed. The gate can be correct about its input while the system is wrong about the customer. That synthetic example is the opening of The…
H
Hello, I am asking about a repeated failure connecting to a remote MCP server, and whether this is a known server-side issue or if there is a workaround. Setup (generic details only): - Remote MCP server: https://mcp.desktopcommander.app/mcp - OAuth 2.0 with dynamic client registration - Scopes: mcp:tools and desktop:auth - Resource: the server root at the same host, without the /mcp path - Token endpoint: Desktop Commander Remote MCP - beta - Token auth: none - PKCE: S256 Failure, repeated:…
H
Subject: Request to register GKB-1000 as a Hugging Face Benchmark Hello Hugging Face team, I would like to request the registration of my dataset, General Knowledge Benchmark (GKB-1000), as an official benchmark on the Hugging Face Hub. Dataset: Lelonthecodeur/General-knowledge-benchmark-for-llm at main GKB-1000 contains 1,000 questions across five categories: Logical and mathematical reasoning Scientific and technical knowledge Language and communication Agentic and adaptive reasoning Critical…