Join the conversation

Join the community of Machine Learners and AI enthusiasts.

Sign Up

All HF Hub posts

SeaWolf-AIΒ 
posted an update about 23 hours ago
view post
Post
4575
πŸš€ Introducing MARL β€” Runtime Middleware That Reduces LLM Hallucination Without Fine-Tuning

Now available on PyPI Β· GitHub Β· ClawHub Β· HuggingFace
AI models sense they could be wrong, but they can't actually fix what's broken.

πŸ€— Live A/B test: VIDraft/MARL

We evaluated 9 SOTA models (GPT-5.2, Claude Opus 4.6, Gemini 3 Pro, etc.) across 1,800 assessments in FINAL Bench and found a 39.2%p gap between "recognizing potential errors (MA=0.694)" and "actually finding and fixing them (ER=0.302)."

MARL (Model-Agnostic Runtime Middleware for LLMs) was built to close this metacognitive gap. It decomposes a single LLM call into a 5-stage expert pipeline (Hypothesis β†’ Solver β†’ Auditor β†’ Adversarial Verifier β†’ Synthesizer), transforming "answer in one shot" into "think, doubt, correct, and rewrite."

No weight modification β€” works instantly with GPT-5.4, Claude, Gemini, Llama, or any OpenAI API-compatible LLM by changing one line: base_url. Ships with 9 domain-specific emergence engines (invention, pharma, genomics, chemistry, ecology, law, and more β€” 5,538 expert data items) activated by a simple tag like model="gpt-5.4::pharma".

pip install marl-middleware

MARL is also officially registered on ClawHub, the skill marketplace of OpenClaw β€” an AI agent platform with 260K+ developers and 3,200+ skills. It's the first middleware in the Reasoning Enhancement category. One command β€” clawhub install marl-middleware β€” gives your AI agent a metacognition upgrade.

πŸ“ Technical deep dive: https://huggingface.co/blog/FINAL-Bench/marl-middleware
πŸ“¦ PyPI: https://pypi.org/project/marl-middleware/
πŸ™ GitHub: https://github.com/Vidraft/MARL
πŸ¦€ ClawHub: https://clawhub.ai/Cutechicken99/marl-middleware

#MARL #LLM #Hallucination #Metacognition #MultiAgent #AIMiddleware #FINALBench #OpenClaw #ClawHub #PyPI #AGI #HuggingFace #ReasoningAI #SelfCorrection #GlassBoxAI
LyteΒ 
posted an update 2 days ago
view post
Post
5287
Introducing Nanochat Moroccan

Nanochat Moroccan is the first language model family built specifically for Moroccan Darija.

This project brings together a small family of models and datasets centered on Darija, with the goal of building something genuinely useful for a language that is still underserved in AI.

1. Models

- KandirResearch/Nanochat-Moroccan-Base-0.7B
- KandirResearch/Nanochat-Moroccan-Instruct-0.7B-pt-raw
- KandirResearch/Nanochat-Moroccan-Instruct-0.7B

2. Data

- Lyte/darija-pretraining-corpus
- Lyte/darija-pretraining-corpus-nanochat
- Lyte/Moroccan-Darija-Instruct-573K
- GemMaroc/TULU-3-50k-darija-english

3. Collection

- https://huggingface.co/collections/KandirResearch/nanochat-the-first-moroccan-darija-language-model-family

Moroccan Darija is spoken by millions of people, yet it remains underrepresented in language technology. Nanochat Moroccan is a step toward building tools that take the language seriously.

You are welcome to try it and chat with it here:
Lyte/Nanochat-Moroccco-Instruct
SeaWolf-AIΒ 
posted an update 2 days ago
view post
Post
5872
ALL Bench Leaderboard β€” Structural Problems in AI Benchmarking and the Case for Unified Evaluation

FINAL-Bench/all-bench-leaderboard

The AI benchmark ecosystem has three structural problems. Major benchmarks like MMLU have surpassed 90%, losing discriminative power. Most leaderboards publish unverified self-reported scores β€” our cross-verification found Claude Opus 4.6's ARC-AGI-2 listed as 37.6% (actual: 68.8%), Gemini 3.1 Pro as 88.1% (actual: 77.1%). OpenAI's own audit confirmed 59.4% of SWE-bench Verified tasks are defective, yet it remains widely used.

ALL Bench addresses this by comparing 91 models across 6 modalities (LLM Β· VLM Β· Agent Β· Image Β· Video Β· Music) with 3-tier confidence badges (βœ“βœ“ cross-verified Β· βœ“ single-source Β· ~ self-reported). Composite scoring uses a 5-Axis Framework and replaces SWE-Verified with contamination-resistant LiveCodeBench.

Key finding: metacognition is the largest blind spot. FINAL Bench shows Error Recovery explains 94.8% of self-correction variance, yet only 9 of 42 models are even measured. The 9.2-point spread (Kimi K2.5: 68.71 β†’ rank 9: 59.5) is 3Γ— the GPQA top-model spread, suggesting metacognition may be the single biggest differentiator among frontier models today.

VLM cross-verification revealed rank reversals β€” Claude Opus 4.6 leads MMMU-Pro (85.1%) while Gemini 3 Flash leads MMMU (87.6%), producing contradictory rankings between the two benchmarks.

πŸ“Š Article: https://huggingface.co/blog/FINAL-Bench/all-bench
πŸ“¦ Dataset: FINAL-Bench/ALL-Bench-Leaderboard
⚑ GitHub: https://github.com/final-bench/ALL-Bench-Leaderboard
πŸ† Leaderboard: FINAL-Bench/all-bench-leaderboard
🧬 FINAL Bench: FINAL-Bench/Metacognitive
prithivMLmodsΒ 
posted an update 3 days ago
view post
Post
4022
The Qwen3.5 Multimodal Understanding Demo, powered by Qwen3.5-2B, is now available on HF Spaces! It is a lightweight model designed for fast image and video reasoning. Built with Gradio, the demo showcases Image QA, Video QA, object detection, and 2D point tracking, along with real-time token streaming.

πŸ€— Demo: prithivMLmods/Qwen-3.5-HF-Demo
βœ… Collection: https://huggingface.co/collections/prithivMLmods/multimodal-implementations
πŸ”— Qwen3.5-2B: Qwen/Qwen3.5-2B

To learn more, visit the app page or the respective model pages.
ReubencfΒ 
posted an update 3 days ago
view post
Post
2543
πŸš€ I am thrilled to announce the release of a new Konkani LLM!

We've seen some fantastic results for both translation and transliteration tasks, and I'm excited to share this progress with the community.

πŸ“– Read the launch article and see the results: https://huggingface.co/blog/Reubencf/konkani-llm
πŸ€– Explore the model and collection:
konkani


I would love to hear your feedback or see what you build with it! #Konkani #LLM #NLP #HuggingFace #IndicNLP #Konkani
codelionΒ 
posted an update 1 day ago
view post
Post
382
Scaling Pedagogical Pre-training to 10 Billion Tokens

New blog post exploring what happens when you take optimal data mixing insights and scale up the data generation itself.

We built Sutra, a multi-stage framework for generating pedagogical pre-training data guided by a knowledge graph of ~2,000 concepts across 9 domains. The pipeline includes structured content generation, six-dimension quality evaluation, diversity management across 20 content styles, and a cleaning stage to prevent collapse.

The result is codelion/sutra-10B, a 10.2 billion token pedagogical dataset with rich metadata (domain, complexity, prerequisites, quality scores) on every entry.

We trained codelion/SmolLM2-70M on it for 3 full epochs (30.6B tokens) on a single A10 GPU in ~78 hours.

Key finding: perplexity kept improving across epochs, but benchmark gains plateaued fast. At 70M parameters, the model hits a representational ceiling that more data alone can't break through.

Full writeup with comparisons against 7 other datasets, detailed benchmark breakdowns, and connections to recent work on synthetic data scaling, curriculum learning, and data mixing laws: https://huggingface.co/blog/codelion/scaling-pedagogical-pretraining-10-billion-tokens

All datasets at multiple scales (10M, 100M, 1B, 10B) plus seed concepts and an SFT variant are in the Sutra Pedagogical Datasets collection.
AINovice2005Β 
posted an update about 13 hours ago
DavidAUΒ 
posted an update 1 day ago
view post
Post
596
21 Qwen 3.5 Fine Tunes (thinking and instruct) ; reg and uncensored (2B to 27B) exceed benchmarks, and work better than org models.

All are bench marked against org model.
Many exceed all benchmarks of org model.
Claude, GLM, Gemini and other distills.
Thinking AND dedicated Instruct versions.

Core goal: Increase benchmarks, and address long thinking blocks.

Highlights:

9B and 27B instruct "Claude" versions hit 624 and 675 on the "ARC-C" (hard challenge).

Thinking fine tunes exceed org model performance (in thinking mode).

In many cases there is a drastic reduction in thinking block size.

9B Claude Heretic Uncensored, GGUF :
-Neo, Code Imatrix (duel imatrix)
- Updated Jinja template
- Custom tensor enhancements.

DavidAU/Qwen3.5-9B-Claude-4.6-OS-Auto-Variable-HERETIC-UNCENSORED-THINKING-MAX-NEOCODE-Imatrix-GGUF

COLLECTION [21 models]:
https://huggingface.co/collections/DavidAU/qwen-35-08-2-4-9-27-35b-regular-uncensored
sweatSmileΒ 
posted an update 1 day ago
view post
Post
239
Just published a hands-on guide on building a Kubernetes cluster from scratch on AWS EC2 using kubeadm, no managed services, no shortcuts.

If you want to truly understand how the control plane and workers communicate, how pod networking works with Flannel, and how to lock down access with security groups ,then this is the kind of exercise that makes it click.

The guide covers a full 3-node setup (1 control plane + 2 workers) on Amazon Linux 2023, from instance provisioning all the way to deploying your first workload.



Read it here πŸ‘‰ https://www.amitchoubey.dev/posts/kubernetes-cluster-aws-ec2-kubeadm/
AINovice2005Β 
posted an update 2 days ago
view post
Post
310
Just published my first cuda kernel, inspired by Sage Attention. Feel free to try it out ☺️

AINovice2005/attention-int8
  • 2 replies
Β·