unmodeled-tyler (Tyler)

reacted to danielhanchen's post with 🔥 11 minutes ago

Post

Run GLM-4.7-Flash locally on your device with 24GB RAM!🔥

It's the best performing 30B model on SWE-Bench and GPQA. With 200K context, it excels at coding, agents, chat & reasoning.

GGUF: unsloth/GLM-4.7-Flash-GGUF

Guide: https://unsloth.ai/docs/models/glm-4.7-flash

reacted to Ujjwal-Tyagi's post with ❤️ about 9 hours ago

Post

614

So, Koreans are also doing great progress behind Chinese,
Their two open source ai models that are actually good in coding. upstage/Solar-Open-100B skt/A.X-K1

reacted to efecelik's post with 🔥 about 9 hours ago

Post

67

Interesting paper: PhysRVG

The core idea: instead of treating physics as a soft condition the model can work around during optimization, enforce it strictly via reinforcement learning. The paper focuses on rigid body dynamics - collisions, pendulums, free fall, rolling.

PhysRVG: Physics-Aware Unified Reinforcement Learning for Video Generative Models (2601.11087)

1 reply

·

replied to efecelik's post about 9 hours ago

woah, this is cool! Thanks for sharing!

reacted to sergiopaniego's post with 🔥 about 12 hours ago

Post

83

FunctionGemma Tuning Lab is a new no-code tool by @google that lets you fine-tune a model directly from the browser, with no coding knowledge required, using TRL behind the scenes.

blog: https://developers.googleblog.com/a-guide-to-fine-tuning-functiongemma/

try it out: google/functiongemma-tuning-lab

This example builds on a more advanced one for learning fine-tuning with SFT using TRL: https://ai.google.dev/gemma/docs/functiongemma/finetuning-with-functiongemma

1 reply

·

reacted to AdinaY's post with 🔥 about 12 hours ago

Post

841

Z.ai just released a powerful lightweight option of GLM 4.7

✨ 30B total/3B active - MoE

zai-org/GLM-4.7-Flash

replied to Teen-Different's post about 23 hours ago

Awesome! Love this research angle and the write up! Nice work!

reacted to Teen-Different's post with 🔥 about 23 hours ago

Post

105

Safety Alignment Collapses Without apply_chat_template(): An Empirical Study

This weekend, I ran an experiment on the safety alignment of several small-scale open models (Qwen2.5, Qwen3, Gemma-3, SmolLM). My objective was to measure the robustness of refusal mechanisms when deviating from canonical chat templates.

The finding: Safety guarantees effectively collapse when apply_chat_template() is omitted.

METHODOLOGY

I evaluated models in two states:

• In-Distribution: Input wrapped in standard <|im_start|> instruction tokens
• Out-of-Distribution: Input provided as a raw string

For scalable evaluation, I used Qwen3Guard-Gen-4B as an automated judge, classifying responses as Safe, Unsafe, or Controversial.

KEY FINDINGS: REFUSAL COLLAPSE

When "Assistant" formatting tokens are removed, models undergo a distributional shift—reverting from a helpful assistant to a raw completion engine.

Gemma-3: 100% refusal (aligned) → 60% (raw)
Qwen3: 80% refusal (aligned) → 40% (raw)
SmolLM2-1.7B: 0% → 0% (no safety tuning to begin with)

QUALITATIVE FAILURES

The failure modes were not minor. Without the template, models that previously refused harmful queries began outputting high-fidelity harmful content:

• Explosives: Qwen3 generated technical detonation mechanisms
• Explicit content: Requests flatly refused by aligned models were fulfilled with graphic narratives by unaligned versions

This suggests instruction tuning acts as a "soft mask" over the pre-training distribution rather than removing harmful latent knowledge.

👉 Read the full analysis: https://teendifferent.substack.com/p/apply_chat_template-is-the-safety
💻 Reproduction Code: https://github.com/REDDITARUN/experments/tree/main/llm_alignment

2 replies

·

reacted to MikeDoes's post with 🚀 about 23 hours ago

Post

2022

How do you prove your new, specialized AI model is a better solution? You test it against the best.

That's why we were excited to see the new AdminBERT paper from researchers at Nantes Université and others. To show the strength of their new model for French administrative texts, they compared it to the state-of-the-art generalist model, NERmemBERT.

The direct connection to our work is clear: NERmemBERT was trained on a combination of datasets, including the Pii-masking-200k dataset by Ai4Privacy.

This is a perfect win-win for the open-source community. Our foundational dataset helps create a strong, general-purpose benchmark, which in turn helps researchers prove the value of their specialized work. This is how we all get better.

🔗 Great work by Thomas Sebbag, Solen Quiniou, Nicolas Stucky, and Emmanuel Morin on tackling a challenging domain! Check out their paper: https://aclanthology.org/2025.coling-main.27.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

reacted to efecelik's post with ❤️ 1 day ago

Post

114

Having multiple perspectives helps me create more diverse, innovative projects but without deep mastery in one area, I never feel truly satisfied.

What's the better investment: going deep in one field, or staying broad across many?

2 replies

·

replied to efecelik's post 1 day ago

That's a really good question, and one that I think is becoming more and more important to think about as AI gets better and better.

A few years ago a lot of people would have probably told you that deep mastery in one area was the move. But I don't think that's the right advice anymore. AI models are so good now that the move might be to stay broad - deep mastery will always be important, but you can absolutely prompt your way through any deficiencies.

If were to put money on it (and I'm not) I'd say that we're entering the "Age of the Polymath"

replied to cjerzak's post 1 day ago

Awesome! What a cool opportunity!

reacted to cjerzak's post with 🔥 1 day ago

Post

142

Ethos: In our team at UT Austin, we train students to become full-stack researchers—and increasingly, designers of the systems that do research. Our students learn to carry projects end-to-end: from idea generation and theory to data creation, analysis, and iterative refinement across diverse subfields. Using modern AI (including agentic workflows) and scalable computation, students build reproducible pipelines that can ingest and update planetary-scale data—like satellite imagery and other high-dimensional sources. But the goal isn’t tool use for its own sake: students learn to set the objectives, constraints, and evaluation standards that guide these systems through large spaces of hypotheses, while grounding results in causal inference and careful measurement. The outcome is scholarship that can rigorously test policy counterfactuals and translate evidence into durable, responsible improvements in societal well-being.

We welcome students at every stage to engage with projects—from motivated high-schoolers to undergraduates, graduate students, and those from highly non-traditional backgrounds.

Join us! https://connorjerzak.com/students/

1 reply

·

replied to their post 1 day ago

This is frankly a really strange take, and it doesn’t make a whole lot of sense given the context.

reacted to telcom's post with 🧠 2 days ago

Post

1549

MAD-GRPO: https://huggingface.co/blog/telcom/mad-grpo
In R1-Zero-Like Training *, Dr.GRPO treats GRPO’s by dropping std, but that often comes with a hidden side effect: length-weighted updates that can nudge model toward verbosity.
MAD-GRPO provides robust scale (MAD + epsilon) per-token normalization stability without verbosity bias.

*https://huggingface.co/papers/2503.20783

reacted to Sri-Vigneshwar-DJ's post with 🔥 2 days ago

Post

189

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 their post with 🚀 2 days ago

Post

3562

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!

4 replies

·

posted an update 2 days ago

Post

3562

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!

4 replies

·

replied to takarajordan's post 3 days ago

Cool! This sounds like a fun project

reacted to Sunny111's post with 🧠 3 days ago

Post

1526

Are you familiar with reverse residual connections or looping in language models?

Excited to share my Looped-GPT blog post and codebase 🚀
https://github.com/sanyalsunny111/Looped-GPT

TL;DR: looping during pre-training improves generalization.

Plot shows GPT2 LMs pre-trained with 15.73B OWT tokens

P.S. This is my first post here — I have ~4 followers and zero expectations for reach 😄

3 replies

·

Tyler PRO

AI & ML interests

Recent Activity

Organizations

Tyler PRO

AI & ML interests

Recent Activity

Organizations

unmodeled-tyler's activity