Sri-Vigneshwar-DJ (DJ Sri Vigneshwar)

reacted to YerbaPage's post with 🔥 about 6 hours ago

Post

380

🔥 SWE-Pruner can save up to 40% of your Claude Code cost without sacrificing performance. Try it out!

📚 SWE-Pruner: Self-Adaptive Context Pruning for Coding Agents (2601.16746)
💻 https://github.com/Ayanami1314/swe-pruner

Drawing inspiration from how human programmers “selectively skim” source code during development and debugging, SWE-Pruner performs task-aware adaptive pruning for long contexts.

reacted to di-zhang-fdu's post with 👀 about 6 hours ago

Post

365

From Pointers to Footnotes: Why Agents will be “Reference-First”

https://medium.com/@di-zhang-fdu/from-pointers-to-footnotes-why-agents-will-be-reference-first-88d0e9730e37

posted an update about 6 hours ago

Post

49

🏙️ Hugging Face Community Post
Title: 🧬 Experimenting with "Dynamic Chaos" in Tamil SLMs

Hi everyone! I just published a new experimental study on Small Language Model (SLM) resilience.

I took the Qwen2.5-0.5B model and put it through a "Chaos Phase" to see how much weight data a tiny model can lose before its understanding of classical Tamil grammar breaks.

Key highlights of the study:

Target Data: Fine-tuned on the Thirukkural (1,330 couplets + modern explanations).
The Chaos Step: Applied 20% random weight pruning but implemented "Layer Protection" for the Token Embeddings and LM Head to keep the characters readable.
Compression: 4-bit (Q4_K_M) quantization for extreme efficiency.
Result: A surrealist classical Tamil model that is ultra-light (~300MB) and ultra-fast!

Check out the model and the experiment logic here: Sri-Vigneshwar-DJ/qwen-tamil-chaos-v1

reacted to alexnasa's post with 👀 8 days ago

Post

1164

Now also with LTX-2 Detailer (Lightricks/LTX-2-19b-IC-LoRA-Detailer)

Try it out: alexnasa/ltx-2-TURBO

replied to davidmezzetti's post 8 days ago

The train dataset comes from the parent model. Mentioning that would be helpful... @davidmezzetti

reacted to kanaria007's post with ❤️ 8 days ago

Post

2486

✅ New Article: *Observations, Under-Observation, and Repair Loops* (v0.1)

Title:
👁️ Observations, Under-Observation, and Repair Loops: The OBS Cookbook for SI-Core
🔗 https://huggingface.co/blog/kanaria007/observations-under-observation

---

Summary:
SI-Core’s rule is simple: *No effectful Jump without PARSED observations.*
This article turns that slogan into an operational design: define *observation units* (sem_type/scope/status/confidence/backing_refs), detect *under-observation* (missing / degraded / biased), and run *repair loops* instead of “jumping in the dark.”

Key clarification: under-observed conditions may still run *read / eval_pre / jump-sandbox*, but must not commit or publish (sandbox: publish_result=false, memory_writes=disabled).

---

Why It Matters:
• Prevents “we had logs, so we had context” failures: *logs ≠ observations* unless typed + contract-checked
• Makes safety real: even PARSED observations should be gated by *coverage/confidence minima* (declared thresholds)
• Turns OBS into something measurable: *SCover_obs + SInt* become “OBS health” and safe-mode triggers
• Links semantic compression to reality: distinguish *missing raw* vs *compression loss*, and fix the right thing

---

What’s Inside:
• A practical observation-status taxonomy: PARSED / DEGRADED / STUB / ESTIMATED / MISSING / REDACTED / INVALID (+ mapping to core status)
• Per-jump *observation contracts* (required sem_types, allowed statuses, age/confidence limits) + explicit fallback actions
• Fallback patterns: *safe-mode / conservative default / sandbox-only / human-in-loop*
• Repair loops as first-class: ledgered obs.repair_request, PLB proposals, governance review for contract changes
• Testing OBS itself: property tests, chaos drills, golden-diff for observation streams

---

📖 Structured Intelligence Engineering Series
this is the *“how to operate OBS”* layer—so the system can *know when it doesn’t know* and repair over time.

reacted to MonsterMMORPG's post with 👀 8 days ago

Post

3936

Compared Quality and Speed Difference (with CUDA 13 & Sage Attention) of BF16 vs GGUF Q8 vs FP8 Scaled vs NVFP4 for Z Image Turbo, FLUX Dev, FLUX SRPO, FLUX Kontext, FLUX 2 - Full 4K step by step tutorial also published

Full 4K tutorial : https://youtu.be/XDzspWgnzxI

Check above full 4K tutorial to learn more and see uncompressed original quality and size images

It was always wondered how much quality and speed difference exists between BF16, GGUF, FP8 Scaled and NVFP4 precisions. In this tutorial I have compared all these precision and quantization variants for both speed and quality. The results are pretty surprising. Moreover, we have developed and published NVFP4 model quant generator app and FP8 Scaled quant generator apps. The links of the apps are below if you want to use them. Furthermore, upgrading ComfyUI to CUDA 13 with properly compiled libraries is now very much recommended. We have observed some noticeable performance gains with CUDA 13. So for both SwarmUI and ComfyUI solo users, CUDA 13 ComfyUI is now recommended.

5 replies

·

reacted to unmodeled-tyler's post with 👍 8 days ago

Post

3680

NEW MODEL: vanta-research/mox-tiny-1

Excited to share the first model from the new Mox-1 model family from VANTA Research!

Mox-Tiny-1 is finetuned from Llama's 3.1 8B architecture specifically for thinking partnership and high-fidelity AI interaction.

Training domains were carefully selected with this goal in mind, and all datasets used were synthetically generated for the Mox-1 models specifically.

Domains include but not limited to:
- Direct opinions (clear stances on various topics)
- Constructive disagreement (Pushing back on flawed premises)
- Epistemic confidence (calibrated certainty and uncertainty)
- Warmth & presence (Empathetic, human responses)
- Wonder & puzzlement (Intellectual curiosity)

Interaction Examples:

Direct opinions:

User: What's the best programming language?
Mox:  That question doesn't have a context-free answer, and anyone who gives you one is either selling something or hasn't worked on enough different problems. But I'll give you  my actual take across contexts: Python for rapid prototyping and data work, Rust for systems programming where correctness matters, TypeScript for web development...

Warmth & Humor:

User: I've been debugging this issue for six hours. 
Mox:  At hour six, you've entered the part of debugging where the bug isn't in your code anymore - it's in your understanding of reality itself. The semicolon starts looking suspicious. Show me what you're dealing with. Fresh eyes help, and mine haven't been staring at it since lunch.

Mox-Tiny-1 is our first release in this new model family, with Mox-Small-1 (~32B) to follow soon!

5 replies

·

reacted to sagar007's post with 🔥 8 days ago

Post

4112

🚀 I built a Multimodal Vision-Language Model from using Gemma-270M + CLIP!

Just finished training my multimodal model on the full LLaVA-Instruct-150K dataset (157K samples) and wanted to share the results!

🔧 What I Built:
A vision-language model that can understand images and answer questions about them, combining:
- Google Gemma-3-270M (language)
- OpenAI CLIP ViT-Large/14 (vision)
- LoRA fine-tuning for efficiency

📊 Training Stats:
- 157,712 training samples (full LLaVA dataset)
- 3 epochs on A100 40GB
- ~9 hours training time
- Final loss: 1.333 training / 1.430 validation
- Only 18.6M trainable params (3.4% of 539M total)

📈 sagar007/multigemma
Benchmark Results:
- VQA Accuracy: 53.8%
- Works great for: animal detection, room identification, scene understanding

🔗 **Try it yourself:**
- 🤗 Model: sagar007/multigemma
- 🎮 Demo: https://huggingface.co/spaces/sagar007/Multimodal-Gemma
- 💻 GitHub: https://github.com/sagar431/multimodal-gemma-270m

Built with PyTorch Lightning + MLflow for experiment tracking. Full MLOps pipeline with CI/CD!

Would love to hear your feedback! 🙏

#multimodal #gemma #clip #llava #vision-language #pytorch

9 replies

·

posted an update 8 days ago

Post

283

Performance Marketing meets "Thinking Mode" 🧠

I’m excited to release hawky-ai-Qwen3-0.6B-Marketing-MoT, a specialized SLM designed for deep strategic reasoning in performance marketing.

While small at 0.6B parameters, this model punches way above its weight class by utilizing a Mixture of Thoughts (MoT) framework. It doesn't just give you an answer; it thinks through the logic of Meta Ads scaling, GA4 attribution, and unit economics before providing a strategic recommendation.

Key Features:

Thinking-First: Trained on 1,500+ critical thinking scenarios.
MoT Framework: 5 distinct reasoning styles (Linear, Exploratory, Critical, Deconstructive, Analogical).
SLM Speed: Perfect for low-latency, high-precision marketing audits.
Check it out on Hugging Face: 🔗 Sri-Vigneshwar-DJ/hawky-ai-Qwen3-0.6B-Marketing-MoT

reacted to MikeDoes's post with 👀 14 days ago

Post

2309

Building powerful multilingual AI shouldn't mean sacrificing user privacy.

We're highlighting a solution-oriented report from researchers Sahana Naganandh, Vaibhav V, and Thenmozhi M at Vellore Institute of Technology that investigates this exact challenge. The direct connection to our mission is clear: the paper showcases the PII43K dataset as a privacy-preserving alternative to high-risk, raw multilingual data

The report notes that our dataset, with its structured anonymization, is a "useful option for privacy-centric AI applications." It's always a delight when academic research independently validates our data-first approach to solving real-world privacy problems.

This is how we build a safer AI future together.

🔗 Read the full report here to learn more: https://assets.cureusjournals.com/artifacts/upload/technical_report/pdf/3689/20250724-59151-93w9ar.pdf

🚀 Stay updated on the latest in privacy-preserving AI—follow us on LinkedIn: https://www.linkedin.com/company/ai4privacy/posts/

#OpenSource
#DataPrivacy
#LLM
#Anonymization
#AIsecurity
#HuggingFace
#Ai4Privacy
#Worldslargestopensourceprivacymaskingdataset

1 reply

·

reacted to AdinaY's post with 🔥 14 days ago

Post

2592

Based on 2025 Chinese AI Timeline, here are some interesting takeaways:

✨ DeepSeek cadence: They shipped almost every month! (except Feb 2025)

✨ Qwen trajectory: Not a single “hit” model, but an expanding product line. VL/Math/Coder/Reranker/Embedding/Omni/Next/Image

✨ Multimodal trend: Steadily rising share, shifting from generation to editing + tooling.

✨ Reasoning as a main track: more engineered, system-level reasoning.

✨ From foundation to components: growth in infra models (embeddings, rerankers, OCR, speech) signals a move toward deployable stacks.

✨ Ecosystem broadening: more players beyond the top labs.

Follow for more updates👉

zh-ai-community

2 replies

·

reacted to dhruv3006's post with 🔥 14 days ago

Post

2312

GraphQL support in Voiden

GraphQL has steadily gained popularity among developers and serves as a powerful query language for your API with a server-side runtime to execute those queries.

Unlike traditional REST APIs, GraphQL lets clients specify exactly what data they need, fetching just that data in a single request, and its strongly-typed schema ensures responses follow a well-defined format .

By allowing clients to select only the fields they require, GraphQL eliminates the over-fetching issues common in REST APIs .

Many developers in our community have asked for GraphQL support, so we have now added it in the latest Voiden release.

What does this mean for you?

- Test and document GraphQL queries and mutations side-by-side with your REST and WebSocket APIs.
- Keep everything in one file-centric, Git-native workflow.
- Use reusable blocks and version control for GraphQL requests, just like any other API call.
- Benefit from a unified environment for all your API design, testing, and documentation.

At Voiden, we believe API work should be treated like code, and that includes protocols like GraphQL.

Try the beta here : https://voiden.md/beta

posted an update 15 days ago

Post

2169

Introducing Hawky-AI H1 4B PM: The First Open-Source LLM for Performance Marketing 🎯

Hey HF Community! 👋

Just released the first LLM fine-tuned specifically for Performance Marketing.
What is it?
Gemma 3 4B distilled from Claude Opus 4.5 with expert-level marketing knowledge.
Covers:
📱 Meta Ads (campaign structure, bidding, scaling, creative fatigue)
🔍 Google Ads (Quality Score, Performance Max, lead gen)
📊 Measurement (ROAS vs MER, incrementality, LTV:CAC)
🎨 Creative Strategy (hook rates, A/B testing, funnel creative)
Why we built it:
Generic LLMs say "optimize your targeting" — not helpful. This model gives specific frameworks like "frequency at 4.5 + CTR drop = creative fatigue, here's the fix..."
Technical:

Base: Gemma 3 4B
Method: QLoRA (r=64)
Teacher: Claude Opus 4.5

🔗 Model: Sri-Vigneshwar-DJ/hawky-ai-H1-4b-PM
Built by Hawky.ai

Try it and let us know what you think! 🚀

reacted to DawnC's post with 🔥 16 days ago

Post

2918

VividFlow: AI Image Enhancement & Video Generation 🎬🎨

Bring your images to life with cinematic motion AND create stunning AI backgrounds! VividFlow combines professional-grade video generation with intelligent background replacement in one streamlined platform.

🎭 Dual Creative Powers
Transform any static image into high-quality dynamic videos with smooth, natural motion ranging from 0.5 to 5 seconds. Choose from curated motion templates across 8 categories designed for portraits, products, landscapes, and artistic content. Create photorealistic backgrounds by selecting from 24 professionally crafted scene presets spanning studios, natural environments, urban settings, and artistic atmospheres...etc.

⚡ Optimized Performance
Video generation currently completes in 4-5 minutes with active optimization underway to dramatically reduce processing time. Background replacement finishes in 30-40 seconds after initial loading. The independent dual-tab design ensures smooth workflow without performance conflicts.

🎯 Complete Creative Control
Achieve perfectly consistent results with seed-based reproducibility and adjustable duration for video generation. Background creation offers flexible composition modes, precision edge softening for challenging subjects, and instant mask preview for quality verification.

📈 Continuous Innovation
Ongoing optimization targets significantly faster video generation through advanced model preparation. Future enhancements include expanded template libraries, batch processing capabilities, and industry-specific presets shaped by community feedback.

👉 Try it now: DawnC/VividFlow

Support development with a ❤️ — your engagement shapes future priorities!
#AI #ImageToVideo #BackgroundGeneration #VideoGeneration

2 replies

·

reacted to SkorczyByk's post with 👍 16 days ago

Post

2314

Potrzebuję modelu do analizy długich dokumentów technicznych, z wzorami, tabelami, wykresami, formułami matematycznymi.

2 replies

·

reacted to kanaria007's post with 👀 16 days ago

Post

494

✅ New Article: *Measuring What Matters in Learning* (v0.1)

Title:
📏 Measuring What Matters in Learning: GCS and Metrics for Support Systems
🔗 https://huggingface.co/blog/kanaria007/measuring-what-matters-in-learning

---

Summary:
Most “AI for education” metrics measure *grades, time-on-task, and engagement*.
That’s not enough for *support systems* (tutors, developmental assistants, social-skills coaches), where the real failure mode is: *the score goes up while the learner breaks*.

This guide reframes learning evaluation as *multi-goal contribution*, tracked as a *GCS vector* (mastery, retention, wellbeing/load, self-efficacy, autonomy, fairness, safety) — and shows how to operationalize it without falling into classic metric traps.

> If you can’t measure wellbeing, fairness, and safety,
> you’re not measuring learning — you’re measuring extraction.

---

Why It Matters:
• Moves beyond “grading” into *support metrics* designed for real learners
• Makes *wellbeing, autonomy, fairness, and safety* first-class (not afterthoughts)
• Separates *daily ops metrics* vs *research evaluation* vs *governance/safety*
• Turns “explainability” into *answerable questions* (“why this intervention, now?”)

---

What’s Inside:
• A practical *GCS vector* for learning & developmental support
• How core metrics translate into education contexts (plan consistency, trace coverage, rollback health)
• A tiered metric taxonomy: *Ops / Research / Safety*
• Parent-facing views that avoid shaming, leaderboards, and over-monitoring
• Pitfalls and failure patterns: “optimize test scores”, “maximize engagement”, “ignore fairness”, etc.

---

📖 Structured Intelligence Engineering Series
Formal contracts live in the evaluation/spec documents; this is the *how-to-think / how-to-use* layer.

reacted to mike-ravkine's post with 🔥 16 days ago

Post

1454

In case you missed it, I had some fun playing with allura-forge/Llama-3.3-8B-Instruct over the holidays - its not every day that we get a new "old llama".

https://huggingface.co/blog/mike-ravkine/new-old-llamas

posted an update 18 days ago

Post

1373

🦅 Introducing Hawky AI H1 Mini 4B: A Domain-Specific Model for Performance Marketing

Hey HuggingFace community! 👋

We're excited to share our first open-source release: **Hawky AI H1 Mini 4B Experimental** - a Gemma 3 4B model fine-tuned specifically for Meta advertising and performance marketing strategy.

🎯 Why We Built This

At [Hawky.ai](https://hawky.ai), we build AI-powered creative intelligence tools for performance marketers. We work with major agencies (WPP, Madison, GroupM) and brands (TVS Motors, Tanishq, Bajaj Finserv) on campaign optimization.

We wanted to explore: Can a small, domain-specific model provide expert-level guidance on performance marketing?

Specifically, we focused on Meta's Andromeda algorithm - the AI system that now powers ad delivery across Facebook and Instagram. Understanding Andromeda is crucial for modern media buying, but the knowledge is scattered and constantly evolving.

🧠 What Makes This Different

Chain-of-Thought Reasoning
The model doesn't just answer - it **thinks through problems** step-by-step:

Sri-Vigneshwar-DJ/hawky-ai-h1-mini-4b-experimental

reacted to JLouisBiz's post with 👀 18 days ago

Post

1875

I’ve built two Firefox extensions for my personal workflow:

1. **Quick Edit in Emacs**
I manage over 3,500 web pages locally. With this extension, I can now click anywhere on a webpage and instantly jump into Emacs to edit the exact page (or annotate any other page I'm working on).

2. **Describe Images (and soon Videos) on the Web**
Using the right-click menu, I can generate descriptions for images I come across online. These descriptions are stored and reused for my own image collections or web pages. I’m planning to add the same functionality for videos soon.

What makes this possible is running LLMs locally on my own machine — I’ve been experimenting with models like **Mistral Vibe** and others. This lets me automate description generation and text processing entirely offline, keeping everything fast, private, and fully under my control.

DJ Sri Vigneshwar

AI & ML interests

Recent Activity

Organizations

DJ Sri Vigneshwar

AI & ML interests

Recent Activity

Organizations

Sri-Vigneshwar-DJ's activity