unmodeled-tyler (Tyler)

reacted to paulpham157's post with 🧠 about 4 hours ago

Post

59

Two things to know right before starting:

- Learn Git. Git is a great versioning tool, even when working alone. It's also essential when working in a team. Don't make excuses that you only do DL and can't do software development.
😞 Don't create files like:
main_backup_1.py
main_backup_2.py
main_backup_3.py
anymore...
(It sounds ridiculous, but I've actually seen some people do that... weren't students.)

- Try to keep everything stable. Imagine encountering errors during a demo. Keep the code clean so it runs smoothly and is maintainable. Always minimize DevOps steps to ensure quick reboot (this can be covered by some platforms; thanks to Hugging Face for making it easy and providing a basic infrastructure that most people can access almost for free). 🤤

reacted to Smooke's post with ❤️ about 4 hours ago

Post

47

New

HackerNoon Post: The Words of Interest Benchmark Test For Matching an LLM to Your Interests https://hackernoon.com/the-words-of-interest-benchmark-test-for-matching-an-llm-to-your-interests

By picking individual words instead phrases or paraphrases or passages, this test bypasses plot summaries (which are everywhere regurgitating themselves online) and focuses on the author's words. It reveals whether an AI has truly "absorbed" the specific texture of a book or is simply echoing the general internet consensus.

reacted to phronos-research's post with 👀 about 4 hours ago

Post

23

Can we measure how AI interaction reshapes human cognition? We built two semantic association instruments that pit humans against Claude Haiku—testing divergent thinking and communicability under constraint. Try the instruments and contribute to the dataset: https://instruments.phronos.org/ins-001/

Explanation here: https://phronos.org/dispatches/semantic-cartography

reacted to nyuuzyou's post with 🔥 about 4 hours ago

Post

26

🏛️ Google Code Archive Dataset - nyuuzyou/google-code-archive

Expanding beyond the modern code series, this release presents a massive historical snapshot from the Google Code Archive. This dataset captures the open-source landscape from 2006 to 2016, offering a unique time capsule of software development patterns during the era before GitHub's dominance.

Key Stats:

- 65,825,565 files from 488,618 repositories
- 47 GB compressed Parquet storage
- 454 programming languages (Heavily featuring Java, PHP, and C++)
- Extensive quality filtering (excluding vendor code and build artifacts)
- Rich historical metadata: original repo names, file paths, and era-specific licenses

This is one of those releases that I'm most interested in getting feedback on. Would you like to see more old code datasets?

reacted to their post with 🚀 about 4 hours ago

Post

25

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

Mox-Small-1 has landed on the Hub!

Finetuned from the fantastic Olmo3.1 32B architecture by AllenAI, Mox-Small-1 was trained using the same datasets and methodology as Mox-Tiny-1, making this model our second addition to the Mox-1 family of models.

Mox-1 is designed to prioritize clarity, honesty, and genuine utility over blind agreement. These models are perfect for when you want to be challenged in a constructive, helpful way.

By utilizing Olmo3.1 32B's architecture, Mox-Small-1 brings greater conversational depth and reasoning quality to the Mox-1 model family. Check it out!

posted an update about 4 hours ago

Post

25

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

Mox-Small-1 has landed on the Hub!

Finetuned from the fantastic Olmo3.1 32B architecture by AllenAI, Mox-Small-1 was trained using the same datasets and methodology as Mox-Tiny-1, making this model our second addition to the Mox-1 family of models.

Mox-1 is designed to prioritize clarity, honesty, and genuine utility over blind agreement. These models are perfect for when you want to be challenged in a constructive, helpful way.

By utilizing Olmo3.1 32B's architecture, Mox-Small-1 brings greater conversational depth and reasoning quality to the Mox-1 model family. Check it out!

replied to ZomiLanguage's post about 15 hours ago

Cool project! Good luck! 👍

reacted to ZomiLanguage's post with 🔥 about 15 hours ago

Post

630

🧠🌍 Zomi Language AI — Community-Driven, Open-Source

![Zomi Language AI – From Community to Model]

The **Zomi language** carries identity, faith, and history for its people, yet it remains underrepresented in modern AI systems.

This project introduces a **community-driven, open-source AI translation framework** that enables Zomi to be trained into AI systems **ethically, transparently, and sustainably**—by native speakers, for future generations.

### 🔁 How It Works
🧑‍🤝‍🧑 Community Texts → 📦 Open Datasets → 🤖 AI Training → 📊 Evaluation → 🔁 Community Review

### 🔓 Why Open-Source Matters
- 🤝 Community ownership
- 🕊️ Cultural & faith integrity
- ♻️ Long-term sustainability
- 🔍 Transparent datasets & models

This initiative demonstrates how **low-resource languages can shape the future of inclusive AI** through open collaboration.

> *No language should be digitally invisible.*

**@Zomi Language | fb.com/ZomiLanguage**

### 🏷️ Tags
#OpenSourceAI #LowResourceLanguages #NLP #MachineTranslation #LanguagePreservation #CommunityAI #ZomiLanguage

1 reply

·

reacted to marksverdhei's post with 🔥 about 15 hours ago

Post

734

Inspired by the heroes of day zero quants ( @TheBloke @danielhanchen @shimmyshimmer @bartowski ), I decided to join the race by releasing the first FP8 quant of glm-4.7-flash! Not as easy as i expected, but I'm happy i was still able to have it working within a few hours after the original model was released! Interested in feedback if anyone wants to try it out!

marksverdhei/GLM-4.7-Flash-FP8

Note: If my PR to vLLM isn't merged yet you might have to use my fork. Cheers! 🤗

reacted to danielhanchen's post with 🔥 about 19 hours ago

Post

885

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 ❤️ 1 day ago

Post

1608

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 🔥 1 day ago

Post

788

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)

2 replies

·

replied to efecelik's post 1 day ago

woah, this is cool! Thanks for sharing!

reacted to sergiopaniego's post with 🔥 1 day ago

Post

807

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 🔥 1 day ago

Post

1594

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

✨ 30B total/3B active - MoE

zai-org/GLM-4.7-Flash

1 reply

·

replied to Teen-Different's post 1 day ago

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

reacted to Teen-Different's post with 🔥 1 day ago

Post

783

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 🚀 1 day ago

Post

2719

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 ❤️ 2 days ago

Post

601

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 2 days 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"

Tyler PRO

AI & ML interests

Recent Activity

Organizations

Tyler PRO

AI & ML interests

Recent Activity

Organizations

unmodeled-tyler's activity