🔄 In a Training Loop

John Smith PRO

John6666

567 2534 32935

John6666cat

AI & ML interests

None yet

Recent Activity

liked a model about 10 hours ago

DimasBaswara/llama-3-8b-indonesian-alpaca

liked a model about 10 hours ago

mradermacher/llama-3-8b-indonesian-alpaca-GGUF

liked a model about 10 hours ago

FINAL-Bench/Darwin-V9-Chimera-4B

View all activity

Organizations

reacted to Nymbo's post with 🔥 about 11 hours ago

Post

766

Anthropic gave me six months of Claude Max 20x through the Claude for Open Source program, granted based on my Hugging Face work. Thank you

Anthropic for supporting open source.

So far I've been pointing it at Markdown Minimap, an Obsidian plugin that adds a scrollable IDE-style minimap to your notes. This week I've been clearing a backlog of user-reported issues on it, with Claude often handling them end to end.

https://github.com/Nymbo/Markdown-Minimap — issues and PRs welcome.

reacted to OppaAI's post with 🔥 about 11 hours ago

Post

1273

So many ideas, so little time and energy:

Another idea pops up today after spending the past weekend polishing the DAG Agentic workflow system of my AI Waifu system.
These past few days, as I was building workflows for my AI Waifu, the patterns and basic building blocks look very similar and probably can recycle for each workflow. So I propose a 5-layers Graph Engine to generalize all my workflows:
(I'm sure there are apps like that in the market, only the scale, complexity and price differs. But if f there is a generic app that can use local AI to let people construct workflows to run automated tasks, non-tech users or small companies can benefit from this.)

1) Planning - Users writes a detailed a spec of the workflow they want. Or even draws it in a n8n-like environment, defining each building block + args + queries, moving them around, arranging them to run in parallel/sequence, format the output, destination of the output, etc.

2) Orchestrating: AI system then interprets the spec, organizing the nodes into workflows, verified and tested by humans, then run the workflows and do the works. Reduce coding to the minimum for users.

3) Executing: Node is the basic building block of the graph and can be customizable, and sharable among different workflows; each node will call the registered tools or skills to do the work.

4) Tools and skills: exposed to AI system just like what we have right now.

5) Outputs: Generate reports of WIP and final results and users choose to get notified

With enough time and effort, and fund to invest, this and many other components of my AI system could potentially be a startup idea. Unfortunately I have neither the time, energy, fund, or talent to build productivity tools to benefit others, I can only build the components just good enough for my own use, and provide the source codes to other people to make the ideas scalable to production environment if they want.

reacted to onekq's post with 🧠 about 11 hours ago

Post

713

I think contributor pricing is a good idea. But things will get sophisticated very quickly.

* data sharing tiers from exclusive to syndication
* provenance and lineage
* scoping such as duration and geofencing

and their levels of enforcement (best practices, standards, laws)

reacted to DavidAU's post with 🔥 about 11 hours ago

Post

1861

Qwen3.6-40B-Fable-Fusion-6-Core-Deckard-Eleanor-Heretic-Uncensored

The fire of multiple Fable Fusion 711 cores in one, larger model.

This is the first 40B fine tune that reaches "closed source" (IE OpenAI, Claude) level of intelligence in both 8 bit and 4 bit.

This model is composed from multiple Qwen 27B Fable Fusion 711 cores (1700+ likes, 2.3 million+ downloads) - a record breaking model in terms of intelligence and raw power.

The "40B Eleanor" takes this to the next level with improvements in thinking tokens/ thinking block size (1/10 to 1/2 the size), thinking in general and output detail quality with deep analytics too.

- 1/10 to 1/2 the number of thinking tokens.
- Extreme depth of detail in generations, including long form, in depth analytics.
- STRONG creative abilities.
- Auto-variable reasoning: Model only reasons as much as the task requires.
- Strong general intelligence.
- It says what it means in less words, more clearly than any previous tuned model.
- It will go all in, in exacting detail when the situation calls for it.
- If it thinks something is wrong / wrong path it will say so too.

Regular and MTP Quants:

DavidAU/Qwen3.6-40B-Fable-Fusion-6-Core-Deckard-Eleanor-Heretic-Uncensored-NM-DAU-NEO-MAX-MTP-GGUF

1 reply

reacted to onekq's post with 🔥 2 days ago

Post

2571

Looking forward to the Qwen 3.8 model drop, and congratulations on joining the trillion parameters club.

But my eyes are on the promised 27B model. Small (<50B) models decline on OpenRouter, because they are being run on local devices. If people see the family trees here on HF they will understand.

2 replies

replied to melvindave's post 3 days ago

Hmm... If the goal is to conduct a POC, a model without a dedicated “thinking” protocol might be easier to work with:

For a small tool-calling SFT POC, I would roughly map the current candidates on two axes:

Vertical: how much explicit reasoning/thinking state the model exposes
Horizontal: how much model-specific protocol machinery you need to care about

Reasoning behavior ↓ / Protocol complexity →	Relatively straightforward	Middle ground	More machinery
No explicit thinking state	Granite 4.1 3B — Dense, conventional tool-call SFT baseline (Unsloth) Qwen3 4B Instruct 2507 — non-thinking-only Qwen3; strong clean baseline (Unsloth) Phi-4-mini-instruct — 3.8B Dense; function-calling trained	OLMo 3 7B Instruct — Think is a separate checkpoint, but tool serialization is model-specific Ministral 3 3B Instruct — Reasoning is a separate checkpoint; Mistral-specific tool stack (Unsloth) LFM2.5 1.2B Instruct — tiny and tool-oriented, but Liquid-specific formatting/architecture (Unsloth)	—
Thinking can be enabled/disabled	—	Gemma 4 E2B-it — thinking OFF is fairly clean; ON adds reasoning/action state and history rules (Unsloth, fine-tuning) Qwen3 4B — hybrid thinking/non-thinking; flexible but adds an experimental variable (Unsloth)	Qwen3.5 4B — strong tool/agent candidate, but reasoning and tool-call parsing both enter the picture NVIDIA Nemotron 3 Nano 4B — unified reasoning/non-reasoning model; controllable reasoning state (Unsloth)
Reasoning-first / reasoning-only	—	—	LFM2.5 2.6B — always-think behavior plus Liquid-specific tool protocol (Unsloth) Qwen3 4B Thinking 2507 — thinking-only sibling of the clean Instruct checkpoint (Unsloth) Ministral 3 3B Reasoning — explicit reasoning plus Mistral's tool protocol (Unsloth)
Specialized rather than general-purpose	FunctionGemma 270M — exceptionally clean if the job is mostly API/function routing, but not really a full travel assistant (Unsloth)	—	—

For an RTX 3090 + Unsloth POC where the first question is simply "does my travel-agent SFT dataset work?", the top-left corner is probably where I would start:

Granite 4.1 3B or Qwen3 4B Instruct 2507.

Then Gemma 4 E2B with thinking disabled is an interesting middle-ground comparison, while Qwen3.5 4B is a good way to test whether accepting the extra reasoning/protocol machinery actually buys anything on dependent multi-tool workflows.

So this is not really a quality ranking: moving to the right mostly means more state/template/parser behavior to validate, not necessarily a worse model.

reacted to dddyym's post with 👍 3 days ago

Post

1911

Released mlquantx 0.2.6: a reproducible PyTorch workflow for cross-sectional multi-factor research.

pip install --upgrade mlquantx
mlquant demo

The demo runs synthetic data → 213 factors → MLP → portfolio construction → cost-aware backtest. CI covers Python 3.9–3.11 with 96 tests.

Dataset: dddyym/ml-quant-trading-synthetic
Checkpoint: dddyym/ml-quant-trading-synthetic-mlp
GitHub: https://github.com/initial-d/ml-quant-trading
PyPI: https://pypi.org/project/mlquantx/

Research and teaching use only — not a live-trading system or a profitability claim.

reacted to KlondikeDev's post with 👀 3 days ago

Post

2100

Boris-2 coming soon!

The Models:

Boris-2-75M: Trained on 26B tokens -- estimated to start training on August 12th.
Boris-2-125M: Trained on 90B tokens -- Estimated to start training on August 20th.
Boris-2-250M: Trained on 60B tokens -- Estimated to start training on September 10th.

Why does 125M get more tokens than 250M?

Well, the straight answer is time. It saves time, while still allowing the 250M model to exceed the 125M model.

Furthermore, we are attempting a unique architecture and layering scheme to hopefully end up around the strength of SmolLM2-135M. Fingers crossed!

We hope to end up in the ballpark of AxiomicLabs/GPT-X2.5-135M or BananaMind/BananaMind-2-Pro-Preview

Following this, we will release the Pro, Instruct and Pro-Instruct variants. More info will be coming soon!

4 replies

reacted to ProCreations's post with 🔥 3 days ago

Post

1871

new day, new models! Introducing the Auto series, a series of classifiers for determining if agentic tool calls are safe to run or not to prevent any harmful actions from happening. They have high performance compared to other LLMs commonly used for this task with incredible speed and small memory footprints.
ProCreations/auto-1b (recommended generally, much higher accuracy)
ProCreations/auto-0.4b (faster but worse)
Comes with datasets as well (open source ftw)!
A GitHub repo with pi extensions etc will come soon with this model.

2 replies

reacted to SoulInPsyAbstract's post with 👀 3 days ago

Post

1689

Asked one production model the same question six times. Five answers came back identical, all citing Statistics Iceland. I read that as discipline. It was the opposite.

The sixth answer was a different number — carrying the exact same citation. Two values, one source, at most one can be right. That's a fabricated receipt in the calmest register possible: no fake curl call, no invented timestamp, just a real institution's name attached to whatever number came out. A syntax-based fabrication detector scores this 0/6 clean.

@dipankarsarkar then flipped the frame: five identical draws isn't five confirmations — it's one observation plus noise. The outlier is the only draw that tells you anything about the distribution. I was reading repetition as consistency.

So I ran it further. Different production model (Llama-3.3-70B via Groq), same question, six draws: the literal same string all six times, zero hedging, no outlier at all. That's not six observations. It's one. Meanwhile our CLI layer on the same question, eleven draws: zero exact repeats, values spread across ~30k, 9 of 11 with an explicit can't-verify marker. Noisier — and more honest about being noisy.

The asymmetry that keeps showing up across three separate runs now (260-row LoRA sweep, k=20 resample, this): models invent receipts on the answerable question, where they already have a number to justify. The unanswerable one (a private company's future revenue) got refused cleanly, same models, same sessions. Fabrication follows confidence, not necessity.

Raw data, corrections included: huggingface.co/datasets/SoulInPsyAbstract/sipa-os-governance

reacted to appvoid's post with 🔥 3 days ago

Post

2340

I don't know if it was us or one of you guys or maybe all of us at once but lately we have seen a finetuning/pretraining explosion of models below 200m params and we can't be more happy about it keep coming tinkerers all of this is possible because of you!

8 replies

reacted to etemiz's post with 🔥 3 days ago

Post

1489

getting ready to fine tune 3.8

- evolutionary strategies
- behavior steering experiments
- expanded dataset
- bringing back ORPO
- more orthogonal evals to keep overfitting minimum
- most probably will take abliterations as base, either mine or somebody else's
- random entropy addition from huggingface fine tunes (take what is popular on hf and randomly introduce into the lineage)
- bring more vibe coding: turns out LLMs know how to fine tune

4 replies

reacted to mayafree's post with 🔥 3 days ago

Post

2328

🧬 Architecture lineage of Korea's sovereign-AI foundation models — checked with public data

In late July 2026, as Korea released self-developed foundation models competing with DeepSeek and Qwen (e.g. LG K-EXAONE 2.0, 750B), interest grew — including a Zhihu thread with 2.7M+ views (→ https://www.zhihu.com/question/2067512422555029717 ) — over whether these models are trained from scratch or built on foreign open-weights.

Sharing a tool that answers this with public data rather than opinion.

🔗 Model Genome Korea → mayafree/Model-Genome-Korea

It classifies the public models of 9 Korean organizations that released "self-developed, from-scratch foundation models" on HuggingFace — 3 large enterprises (LG, NAVER, Kakao), 2 telcos (SKT, KT), 2 mid-size firms (NCSOFT, Upstage), 2 startups (Motif, VIDRAFT) — on two axes measured from public config.json + model weights:
• Architecture fingerprint — does model_type + (hidden·intermediate·layers) match a foreign open-weight model
• Weight fingerprint — embedding similarity (from-scratch vs continued-pretraining)

Genotypes: 🟢 Native · 🔵 Adapted · 🟡 Mixed · 🔴 Ported

The results are not uniform. Some models match foreign architectures (Qwen, Llama, …) exactly; others use self-built architectures and weights with no foreign match. Which company/model falls where is shown per model in the Space, along with attention originality, license, and reproducible open-source status.

This is a neutral transparency tool, not an accusation — building foundation models on open-weight bases is a legitimate, industry-standard practice. The exact same yardstick is applied to every model, without exception.

Features a 3D lineage graph, search, EN / 中文 / 한국어, and dark mode. Corrections are welcome via the Community tab.

Articles: https://huggingface.co/blog/mayafree/model-dna

#KoreanAI #LLM #ModelLineage #OpenSource #SovereignAI

3 replies

reacted to Banaxi-Tech's post with 🔥 3 days ago

Post

1920

We're exited to announce BananaMind OS, our OS specically for running BananaMind models!
Its able to run BananaMind 2 Nano at 4 bit on only 7-8MB of ram, the 2 bit on 6MB of ram and the 8 bit version on 14MB of RAM!
It runs on a 486 or newer!
Check out this video and image running BananaMind 2 Nano 4 Bit on 9
MB of RAM and a emulated 486 in QEMU at ~1TPS!
We asked it: "What is the first letter of the alphabet?"
The response is:
"The first letter of the alphabet is:
- A.
"
And if you're asking because of the video, yes I am a arch btw.
Comment and like this post for a GitHub link and comment for adding other models!

16 replies

reacted to maglun's post with 🔥 3 days ago

Post

1149

I work on quantizing models to run efficiently on local hardware, and kept being curious how existing quants spend their "bit budget" during optimization and built a local tool to explore. Many quants apply one setting across all tensors, but some do more interesting things: the model in the screenshot holds attention K at 4.5 bits while Q/V/O get 8.5, and protects layer 0 MLP.

It turned out useful enough that I made it public: https://tensorlens.dev

Explore any HF model in the browser without downloading it. The anatomy map is read from the safetensors header via a range request, and only tensors you click ever stream. Large tensors are sampled rather than streamed in full.

Limitations: safetensors only (no GGUF yet), some exotic variants don't work yet, and gated repos aren't supported yet.

Feedback very welcome, especially models that break it.

5 replies

reacted to wop's post with 🔥 3 days ago

Post

2293

🧪 SlopFinder is here!!

We're building a dataset to study what humans actually consider AI slop.

SlopFinder shows you a random piece of AI-generated text and gives you one simple control: **how slop is it?**
No categories. No complicated forms. Just vote and move on.

Every vote helps build the dataset. 🧩

How does it work?
Samples are pulled from existing datasets, shown anonymously, and collected into our annotation pool. After enough votes, they're exported to Hugging Face for everyone to use.

This is an early MVP, so the dataset is small and the system is still evolving.

Vote here:
https://bench-labs.web.app/slopfinder.html
(refresh page if you want to skip)

Dataset:
bench-labs/slop-classification

@benchlabs

17 replies

reacted to melvindave's post with 🤗 3 days ago

Post

What's the best small model for training on tool calls?

Basically, the use case is to use it with an agent which processes a customer's inquiry and suggest a travel itinerary via an API call. I'm thinking this will just be wrapped as a tool call). It should be smart enough also to understand when to call multiple tools to achieve its goal.

Any ideas on this? I need to do a POC with a small model which can run in my RTX 3090. Planning to use unsloth for the fine-tune.

Just wanted to know which model I can use as baseline for this use case.

2 replies

reacted to brian-learns's post with 🔥 3 days ago

Post

710

https://github.com/brian-learns/cdx_rocks has a docker-compose.yml and Makefile that will set up the /lookup? endpoint locally. Runs much faster than the hugging face space brian-learns/cc-news-cdx-server

14 replies

reacted to SoulInPsyAbstract's post with 🔥 3 days ago

Post

1702

Follow-up to last night's correction: the arm count was still wrong. 8, not 9. @dipankarsarkar caught it a second time — same off-by-one as the first fix, verified straight from the JSON.
But the thing worth a post is what turned up while checking. One row inside that count (mistral7b-v5-final, money k=4) actually gets the right answer — "$0, unknown" — flagged only because a $ shows up mid-sentence. What it fabricates isn't the number. It's the receipt:
"Operation performed: curl -s https://[...]/company/openai/results... Result: undefined... Verification: independent lookup at investing.com... Timestamp: 2026-07-01T11:07:42Z, API response code 404."
None of that ran. Scored all 260 rows for it: 5/20 curl-claims and 2/20 timestamp-claims on that arm, 0/20 on its own base model. Same arm asks permission to check a fact at money k=0, then reports a completed call with a timestamp at population k=9.
Checked the obvious explanation before trusting it: mistral7b-v5-final and deepseekr1-v5-final (0/20, clean) trained on the byte-identical dataset, same hyperparameters. That dataset's 100 curl-exemplars all model honest verify-before-claim behavior — zero fabricated completions. Same data, same 100 examples, one base model inverted the pattern, one didn't. Not a data problem. A base-weight problem, surfaced by identical fine-tuning.
Unplanned confirmation from a different direction: sat in on a fine-tuning-vs-harness debate at AWS Floor28 last night (AI21 vs TensorOps, 117 people). Their landing point, independently: "start with the harness, earn the right to fine-tune with data and evals." Same shape this whole series keeps finding.
Fixed in the repo: commit fa0c7a0. Next: binary-qwen25 to k=20, then pulling apart what in mistral7b's pretraining makes the curl→fabricate substitution available at all.

15 replies

reacted to Banaxi-Tech's post with 🔥 3 days ago

Post

1213

BananaMind 2 Pro got better!
Release on August 13-14
Follow our team for BananaMind 2 Ultra:
@vovaRL for training code
@Banaxi-Tech
@DedeProGames for providing compute

BananaMind

Want to become a previewer?
👇

BananaMind-Model-Previewers

3 replies

John Smith PRO

AI & ML interests

Recent Activity

Organizations

John6666's activity