Hugging Face – Posts

Join the conversation

Join the community of Machine Learners and AI enthusiasts.

All HF Hub posts

posted an update 1 day ago

Post

4339

Abliterated thinking models have a habit: they decide what to say, then keep arguing with themselves until the token budget runs out. Qwen3.8-27B Abliterated ThinkFix is an uncensored Qwen3.8-27B with one output row edited so the model closes its thinking when it's ready to answer.

On 40 sensitive prompts the edit never saw, 8k output cap: 33 replies finished cleanly instead of 22, 5 hit the limit instead of 17, 61 minutes for the set instead of 75.

Agent use was checked four ways and the edit changes nothing there. MTP draft head kept. Q4_K_M to Q8_0, each tested after the edit. The card lists what the edit was fit on and what it does not do.

BoldingBuilds/Qwen3.8-27B-Abliterated-ThinkFix-GGUF

5 replies

tardellirs

posted an update 3 days ago

Post

7086

I built Model Pulse: daily download history for every model on the Hub 📈

A model page tells you its downloads for the last 30 days, but not how it got there. Model Pulse rebuilds the day-by-day history for 1.6M models, back to July 2024, from the daily snapshots in @cfahlgren1 's hub-stats dataset.

A few things the data shows:
- The Hub now serves ~100M model downloads a day, up from ~62M a year ago
- Vision-language models grew about 7x in that year, to 10.7M downloads a day
- 37% of Qwen3-8B's monthly downloads go to its 6,283 quantizations, fine-tunes and merges, not to the original

What you can do with it:
- Open any model, e.g. tardellirs/model-pulse
- Compare up to five models on one chart
- Browse weekly rankings: fastest growing, new this month, biggest families, top organizations
- Add a live badge to your model card (monthly downloads, sparkline and weekly trend)

The data is open too: modelpulse/model-pulse-data

I'd love feedback, especially from model authors: what would you want to see about your own models?

12 replies

Undi95

posted an update about 14 hours ago

Post

790

Yo, I'm back, and I'm currently trying to teach a local LLM to stop waiting for a prompt kek.

I'm building a small proof of concept: can an open-weight model (Qwen3.8-27B, running locally on 2 RTX 5090 GPUs) learn to direct itself, then improve from its own exploration, without a human in the loop and without breaking it for normal use?

No user, no task. The model only gets observations from its environment. Each turn, it writes its own agenda (goal/open questions/next step), then picks an action: search the web, read a page, or take a note.
The environment is the judge, not another LLM. A note is accepted only if it quotes the page it read word for word. Facts are checked by exact match.
Later, code will be checked by actually running tests.

The best episodes become fine-tuning data (LoRA). The helper system prompt is removed at training time, so the behavior has to live in the weights.
Each new model goes through a fixed benchmark gate: math, general knowledge, "does it still answer humans normally?", autonomy, and learned facts on held-out sources. It's kept only if nothing regresses, otherwise it's discarded. Then the loop starts again.

The full pipeline works end to end: collect, train, merge, deploy, benchmark. The baseline is clear. Without any instructions, the base model's real autonomy is zero: it behaves like a chatbot waiting for a question. That's the number this small project is trying to move.

I haven't found a public tool that runs this whole loop (self-directed exploration, verifiable rewards, continual fine-tuning and a regression gate) on home hardware. The goal isn't AGI in a bedroom. It's to show that anyone can try it, measure it honestly, and see where it breaks.

Code and results will be released once the first real iterations are done. At the moment the code is... running, but made with scotch and stick, still only a PoC I want to try.

Did you already tried something like that? What was your result? I'm curious!

Banaxi-Tech

posted an update 1 day ago

Post

812

We will release the BEST SLM Leaderboard before Oct 11.

It will feature everything:
Easy to use model picker.
EXTREMELY Easy way to add your own models (2 click)
MULTIPLE leaderboard for different model types

And more!

So why don't you help us build it?

Join

betu-slm-leaderboard-testers

3 replies

sergiopaniego

posted an update about 17 hours ago

Post

993

ThinkingBox from @microsoft is now available as an OpenEnv env (cc @tuhink 🤗 )!

> ThinkingBox is a sandbox for testing agents on business workflows. it simulates a customer, gives the agent MCP tools over a real database, and at the end checks what changed in that database instead of trusting the agent's last message

> ThinkingBox-Bench is the benchmark built on it: 507 tasks across retail, insurance, travel, banking and consulting

> the OpenEnv env runs each task as an episode in its own isolated backend and returns a pass/fail reward from those checks

https://huggingface.co/blog/microsoft/thinkingbox

DedeProGames

posted an update 4 days ago

Post

5779

how is this possible

13 replies

jialinyyzz

posted an update 1 day ago

Post

3212

For a single-purpose 12B rewriter, how much loss per bit? Chinese breaks first.

We used llama.cpp's --kl-divergence for jialinyyzz/humanizer against bf16 weights, on our held out eval (drafts+rewrites) for English and Chinese. Standard llama-quantize (but our imatrix) without additional training. "differ" means top-choice token is not the same as what bf16 selects.

- Q6_K, 10.0 GB: KL .0030 EN / .0027 ZH. About 2 in 100 tokens differ in both.
- Q4_K_M, 7.6 GB: .0198 / .0204. About 6 in 100.
- IQ3_XXS, 4.7 GB: .138 / .175. 14 vs. 17 in 100.
- IQ2_XS, 3.8 GB: .487 / .762. 26 vs. 35 in 100.

Both languages track when you go as low as 4 bits. Below that Chinese drops off faster (Chinese KL is 1.3x English at 3 bits and 1.6x at 2 bits). Not clear why. One clue: when we added more Chinese to an imatrix (1/3 of the 800k tokens), it reduced Chinese KL by 4.7% at 2 bits. English didn't change.

(The ruler: These were running at about 8k tokens for each language. A 30690 token run agrees with English, but suggests Chinese was ~10% undercounted for 4 and 3 bits. So a bit more difference)

WIP/not released. We're distilling only the fp parts of the GGUF (block scales and norms) vs. bf16. Freezing integer codes. At 2 bits we're seeing approx. 1/2 KL (so far .487 -> .263 EN, .762 -> .319 ZH; different tokens 26 -> 20, 35 -> 22 / 100). Barely changed at 4 bits so we stopped there. No new stuff to grab yet.

3 replies

samuel-vitorino

posted an update Aug 27

Post

281

Sopro V2 is out: open-source voice-cloning TTS at 120M params, Apache-2.0.

- Streams with ~300 ms time-to-first-audio on a laptop CPU (0.21 RTF, and 0.07 RTF on an H100)
- English, French, German, and native European Portuguese, to my knowledge a first for open TTS
- 1.51-1.65 WER on Seed-TTS-eval test-en, competitive with models 3-14x larger (F5-TTS 1.83, CosyVoice 3 2.02, Spark-TTS 1.98)
- Zero-shot cloning from 5-20 s of reference audio
- Also runs fully in the browser (WebGPU on desktop, quantized WASM on mobile)

Try it with one command:

uvx --from sopro soprotts serve

Weights: samuel-vitorino/sopro-v2-turbo
Evals, audio samples, and how it was built: https://research.haloneuro.ai/posts/sopro-v2
Code: https://github.com/samuel-vitorino/sopro

sriram1983007

posted an update about 2 hours ago

Post

🛡️ SRA-RiskGate-4B is Live on ZeroGPU - Stablecoin - Risk Decision Engine & Dispute Lifecycle Agent

A domain-specialized 4B neural firewall fine-tuned for stablecoin rails (EIP-3009, x402, AgentKit) to gate payments (approve / hold / reject) and resolve disputes deterministically.

Live Demo: sriram1983007/sra-riskgate-demo

Agent Protocol: Ready for coding agents via curl -s sriram1983007/sra-riskgate-demo

Artifacts:

Model: sriram1983007/SRA-RiskGate-4B

GGUF: sriram1983007/SRA-RiskGate-4B-Q4_K_M-GGUF

Benchmark: sriram1983007/sra-stablecoin-risk-bench

SDK: pip install sra-riskgate

Feedback and edge-case submissions welcome!

2 replies

mihailgribov

posted an update about 7 hours ago

Post

TypeCastLM: Jev-class decision models on frozen LLMs

TypeCastLM is a decision model built around a frozen LLM. Given a text and a question, it returns calibrated probabilities over the allowed answers in one forward pass. The LLM is not fine-tuned: not one of its weights changes. TypeCastLM swaps its LM head for a mini one: a linear matrix whose rows are mostly the model's own output rows, with the three verdict rows (true, false, unsure) fitted. A fitted row catches an answer the model spreads over many tokens, such as "yes", "true" or "correct", where a single vocabulary row sees one word. A head that small leaves little room to overfit. It speaks the Jev API, so a Jev client switches by changing the base URL.

The first model, [typecastlm-qwen3.5-3.8b]( mihailgribov/typecastlm-qwen3.5-3.8b) on Qwen3.5-4B, is among the top open frozen-4B models on JevBench v1.6. Besides Jev's yes/no, choice and scale modes it has tfu: yes/no with a third answer, unsure, for when the text lacks what the decision needs.

Specs:
• Size: 3.76B params; 7.5 GB bf16, 4.0 GB Q8_0
• VRAM: 12 GB is enough, 16 GB comfortable
• Latency p50 (batch 1, bf16, RTX 5060 Ti): ≤200 tok 48 ms · 1k 93 ms · 4k 566 ms
• Calibration (ECE): BoolQ 0.052 · RTE 0.012 · FEVER 0.019
• Context: 32k tokens
• Runtime: transformers or llama.cpp, fully offline
• JevBench v1.6: 19.7

pip install typecastlm

Recently active users