Hugging Face – Posts

Join the conversation

Join the community of Machine Learners and AI enthusiasts.

All HF Hub posts

posted an update 2 days ago

Post

8337

🧬 Introducing Darwin-9B-NEG — the first model with Native Entropy Gating (NEG)

🔗 Try it now: FINAL-Bench/Darwin-9B-NEG
🔗 Q4 bit : FINAL-Bench/Darwin-9B-MFP4

We're thrilled to release Darwin-9B-NEG, a 9B-parameter reasoning model
that embeds an architecturally-internalised sense of self-confidence directly
into the transformer — our proprietary Native Entropy Gating (NEG) technology.

📊 GPQA Diamond (198 PhD-level questions):

▸ Baseline Darwin-9B (no NEG) → 51.01 %
▸ Pure NEG (greedy · 1× cost) → 63.64 % 🔥 +12.63 %p
▸ + Permutation (4× cost) → 76.26 %
▸ + Ensemble Refinement (~20×) → 84.34 % 🏆

With only 9 billion parameters and 1× inference cost, Pure NEG jumps
+12.63 %p over the same model without NEG. Going all-in with ensemble
refinement pushes it to 84.34 % — surpassing the published Qwen3.5-9B
leaderboard score (81.7 %) by +2.64 %p.

🔬 What makes NEG different from Multi-Turn Iteration (MTI)?

Classical MTI needs 3-8× extra inference passes. NEG instead lives
INSIDE the single decoding loop. Two tiny modules ride with the
transformer: NEG-Head predicts per-token entropy from the last hidden
state, and NEG-Gate conditionally restricts the top-k choice when
confidence is low. The gate activates in only 4.36 % of tokens —
essentially free at inference time.

✨ Key differentiators
• Architecturally internalised — model file *is* the feature
• 1× inference cost (vs. 3-8× for MTI)
• Drop-in with vLLM / SGLang / TGI / transformers — no extra engine
• +12.63 %p reasoning at zero latency overhead
• Single-file deployment, Apache 2.0 licensed

🧬 Lineage
Qwen/Qwen3.5-9B → Darwin-9B-Opus (V7 evolutionary merge) → Darwin-9B-NEG (V8 + NEG training)

#Darwin #NEG #NativeEntropyGating #GPQA #Reasoning #LLM #OpenSource #Apache2

AkimfromParis

posted an update 3 days ago

Post

2406

🌸 𝙊𝙥𝙚𝙣 𝙅𝙖𝙥𝙖𝙣𝙚𝙨𝙚 𝙇𝙇𝙈 𝙇𝙚𝙖𝙙𝙚𝙧𝙗𝙤𝙖𝙧𝙙 𝙑2 𝙤𝙣 𝙃𝙪𝙜𝙜𝙞𝙣𝙜 𝙁𝙖𝙘𝙚 🇯🇵 // 🌸 ハギングフェイス版「 𝗢𝗽𝗲𝗻 𝗝𝗮𝗽𝗮𝗻𝗲𝘀𝗲 𝗟𝗟𝗠 𝗟𝗲𝗮𝗱𝗲𝗿𝗯𝗼𝗮𝗿𝗱 𝗩𝟮 」公開 🇯🇵

I am thrilled to announce the launch of version 2 of the 𝙊𝙥𝙚𝙣 𝙅𝙖𝙥𝙖𝙣𝙚𝙨𝙚 𝙇𝙇𝙈 𝙇𝙚𝙖𝙙𝙚𝙧𝙗𝙤𝙖𝙧𝙙. This initiative is driven by the "Fine-tuning and Evaluation" team, led by Professor Miyao at the The University of Tokyo, under the Research and Development Center for Large Language Models (LLMC) at Japan’s National Institute of Informatics (NII).

𝙎𝙩𝙧𝙖𝙩𝙚𝙜𝙞𝙘 𝙖𝙣𝙙 𝙩𝙚𝙘𝙝𝙣𝙞𝙘𝙖𝙡 𝙪𝙥𝙜𝙧𝙖𝙙𝙚𝙨:
- Our new backend features eight A100 GPUs, enabling the evaluation of open-source models of more than 100B parameters.
- Submissions now require a Hugging Face Hub login to ensure accountability.
- We have added metrics for evaluation time, CO₂ emissions (thx to Code Carbon 🌱 ), alongside reasoning capabilities.

𝘿𝙖𝙩𝙖𝙨𝙚𝙩𝙨 𝙖𝙣𝙙 𝙚𝙫𝙖𝙡𝙪𝙖𝙩𝙞𝙤𝙣 𝙨𝙩𝙖𝙣𝙙𝙖𝙧𝙙𝙨:
- New datasets cover reasoning, mathematics, exams, and instruction following.
- Math evaluations now span from grade-school levels to expert-tier challenges (GSM8K, PolyMath, AIME).
- While integrating English-heavy and multilingual benchmarks (including Humanity’s Last Exam, GPQA, and BBH in both English and Japanese), we continue to prioritize unique Japanese cultural datasets.

llm-jp/open-japanese-llm-leaderboard-v2

どうぞお願い致します！😊

SeanLee97

posted an update 5 days ago

Post

7993

Our lab recently released a paper where we introduce ShadowPEFT, a new Parameter-Efficient Fine-Tuning (PEFT) paradigm tailored for edge computing scenarios.

Unlike traditional approaches such as LoRA and its variants, which inject trainable parameters directly into the weights of Transformer, requiring tight coupling with the backbone.

ShadowPEFT instead enhances the frozen large base model by adding a lightweight, centralized, pretrainable, and detachable Shadow network.
This shadow network operates in parallel with the base model, delivering learned corrections to each decoder layer. Because the shadow module is architecturally decoupled from the backbone, it can be independently trained, stored, and deployed, benefiting edge computing scenarios and edge-cloud collaboration computing.

- HF Paper: ShadowPEFT: Shadow Network for Parameter-Efficient Fine-Tuning (2604.19254)
- GitHub: https://github.com/ShadowLLM/shadow-peft
- HF Collection: https://huggingface.co/collections/shadow-llm/shadow-peft-models

7 replies

Tonic

posted an update 3 days ago

Post

3195

🙋🏻‍♂️ Hey there folks ,

I'm sharing huggingface's largest dataset of annotated statelite images today.

check it out here : NuTonic/sat-image-boundingbox-sft-full

I hope you like it , the idea is to be able to use this with small vision models 🚀

anakin87

posted an update 3 days ago

Post

3185

A small model that struggled against a random opponent now beats GPT-5-mini at tic-tac-toe

I took LiquidAI/LFM2-2.6B and trained it through play.

🧑‍🍳 Here's how:

1️⃣ Build a solid RL env with Verifiers (Prime Intellect)
2️⃣ Generate synthetic data: <200 games sampled from GPT-5-mini playing in the env
3️⃣ SFT warm-up to teach format
4️⃣ Group-based RL (CISPO) against opponents making 20-70% random moves
5️⃣ RL again with stronger opponents (0-25% random moves) + 1.25 temperature to push exploration and shake off suboptimal strategies

Done! Beats GPT-5-mini 🏆

---

🎮 Play against the model: anakin87/LFM2-2.6B-mr-tictactoe

🤗 Model: anakin87/LFM2-2.6B-mr-tictactoe

📚 Walkthrough/course: https://github.com/anakin87/llm-rl-environments-lil-course

🤗 Dataset and checkpoints: https://huggingface.co/collections/anakin87/lfm2-26b-mr-tic-tac-toe

Benedictat

posted an update 3 days ago

Post

3524

Built a WeChat Mini Program in 20 minutes flat with Hy3 Preview + WorkBuddy…

and I didn’t type a single line of code. Not even a semicolon.

This Coding Agent is on steroids. Its comprehension in long back-and-forths is night and day better, and that 256K context window swallows the entire project structure whole.

Tell it what you want, and it actually gets the full picture no confused blank stares from the AI.

And we’re not messing around with dinky little code snippets here. It spits out a fully functional project

app.json, every page’s wxml/wxss/js/json, even Mock data pre-packed. Import it into WeChat Dev Tools and it runs on the first try

Only one tiny visual nitpick, zero logic bugs. Point out the flaw, and it fixes it instantly no new bugs, no passive-aggressive code breaks, no headaches

The entire vibe Tell it your idea → Get a complete working project → Mention a tiny flaw → AI polishes it.

No coding, no endless edits, no soul-crushing debugging that makes you want to throw your laptop. Absolute game-changer

imnotkitty

posted an update 3 days ago

Post

3869

tencent/Hy3-preview is out: an open-weights MoE reasoning model.

✅ 295B total / 21B active / 256K context
✅ Fused fast-and-slow thinking in a single model
✅ First model trained on Hunyuan's rebuilt pretraining + RL infra (Feb → Apr)

Benchmarks:
👉 SWE-Bench Verified, Terminal-Bench 2.0, BrowseComp, WideSearch — competitive results, particularly strong on agentic tool use
👉 Top score on Tsinghua's 2026 Spring math PhD qualifying exam
👉 Strong context-learning and instruction-following on Tencent's CL-bench / CL-bench-Life

More details can be found in my article: https://huggingface.co/blog/imnotkitty/hy3-preview

2 replies

prithivMLmods

posted an update 3 days ago

Post

1052

Now, a collection of various compression schemes for Qwen3.6 and the abliterated version 1 of dense models is available on the Hub. Check it out via the links below. 👇

🔗 Qwen3.6-MoE: https://huggingface.co/collections/prithivMLmods/qwen36-35b-a3b-compressions
🔗 Qwen3.6-27B Compressions: https://huggingface.co/collections/prithivMLmods/qwen36-27b-compressions

🤗 > To learn more, visit the app page or the respective model pages.

qgallouedec

posted an update about 4 hours ago

Post

TRL v1.3 ships day-one training support for Qwen 3.6 🚀

The new Qwen 3.6 family (Qwen/Qwen3.6-27B, Qwen/Qwen3.6-35B-A3B) reuses the Qwen3.5-MoE architecture but ships a slightly different chat template, so we updated the stack end-to-end: new training template with {% generation %} markers, tool-call response schema routing, tiny test models for the VLM matrix.

SFT with assistant-only loss works out of the box:

from trl import SFTConfig, SFTTrainer

trainer = SFTTrainer(
    model="Qwen/Qwen3.6-27B",
    args=SFTConfig(assistant_only_loss=True),
    train_dataset=dataset,
)
trainer.train()

So does GRPO tool-calling — just hand tools=[...] to GRPOTrainer.

v1.3 also brings a new experimental TPO trainer (Triple Preference Optimization), speculative decoding in trl vllm-serve (Qwen3 MTP / Eagle3 drafts), 12 more KTO ↔ DPO alignment PRs (KTO promotion to stable is now in reach), three more {% generation %} chat templates (Gemma/Gemma 2, Phi-3, GLM-4-MoE), and a chunky SFT entropy bug fix.

Full release notes: https://github.com/huggingface/trl/releases/tag/v1.3.0

akhiilll

posted an update about 12 hours ago

Post

Just shipped ClaimSense Adjudication Gym at OpenEnv Hackathon 2026 (Scaler India).

An OpenEnv RL environment for enterprise insurance claims adjudication—the monthly “tool-heavy” workflow real adjusters do: pull policy + claim history, run fraud checks, verify purchase/transactions, then approve / deny / escalate under partial observability with long-horizon credit assignment.

Trained Qwen/Qwen2.5-1.5B-Instruct with:

Rollout evaluation on HF Jobs (A10G) and a random baseline for comparison
Real GRPO weight updates (TRL GRPOTrainer) with LoRA adapters and two independent reward functions (format + env replay)
Headline training evidence:

GRPO run: 80 steps, 640 rollouts, KL rises ~0 → ~0.06 (real weight updates), completion length shrinks (~25 → ~10).
Plots + logs are committed in the Space under runs/.
Live demo + repo + writeup linked below.

🔗 Env (Space URL): akhiilll/claims-env
🧪 Notebook: akhiilll/claims-env
📝 Blog: docs/HF_MINI_BLOG.md in the Space

Recently active users