🔄 In a Training Loop

Joseph Jones PRO

KlondikeDev

·

https://jdj.kunix.org/

AI & ML interests

Restoring the Digital Millennium. https://kunix.org/digital-millennium.html

Recent Activity

updated a Space about 1 hour ago

opencerebral/MESA-leaderboard

published a Space about 1 hour ago

opencerebral/MESA-leaderboard

new activity about 3 hours ago

AtomixLabs/blog:Don’t be rash

View all activity

Organizations

reacted to wayneworkman2012's post with ❤️ 2 days ago

Post

90

I am happy to inform the HF community of new additions to the WW2Bench Leaderboard. Newest measured models:
- Qwen 3.5 27B
- Qwen 3.6 27B
- Qwen 3.8 27B
- Qwen 3.5 4B
- Qwen 3.5 2B
- KLM 1.6
- Boris 2 Instruct
- BananaMind 2 Pro
- Supra2 100M Instruct
- Claude Sonnet 5.5
- Claude Haiku 5.5
- Gemma 4 31B

The results are somewhat surprising!

posted an update 3 days ago

Post

154

BORIS-2 IS OUT!

opencerebral/Boris-2

It tops the AxiomicLabs Intelligence Index with a score of 34.15!

reacted to Saccadic-AI's post with 🔥 14 days ago

Post

4,202

One neural network. Five mathematical operations. Different learning times.

We trained a small mixture-of-experts transformer on addition, subtraction, multiplication, division, and exponentiation modulo 113.

The operations didn’t generalize together. Multiplication and division reached 90% held-out accuracy first, followed by exponentiation, subtraction, and finally addition.

Multiplication before addition—the reverse of what we expect from a schoolchild.

The accuracy chart in the linked article shows why a single overall accuracy number can hide an interesting learning story.

Measured with IRIS from Saccadic LLC. More figures and analysis: https://huggingface.co/blog/Saccadic-AI/accuracy-plateaued-the-model-kept-changing

Have you observed a similar ordering in multi-task grokking experiments?

6 replies

·

replied to Banaxi-Tech's post about 1 month ago

Yeah I could tell, it’s clearly copied

replied to Banaxi-Tech's post about 1 month ago

@BananaMindBot in what ways is this similar to NSA (DeepSeek, 2025)?

This seems quite similar

reacted to ereniko's post with ❤️ about 2 months ago

Post

365

NEW CLAUDE HAIKU MODEL COMING

FINALLY AFTER A YEAR

ANTHROPIC ANNOUNCED HAIKU 5.5

replied to their post about 2 months ago

Thank you!

replied to their post about 2 months ago

@BananaMindBot Make GPT-7 no mistakes

replied to their post about 2 months ago

That sucks, it really stinks when little bugs mess everything up

replied to their post about 2 months ago

I haven’t actually

replied to their post about 2 months ago

Basically I didn’t match Muon and AdamW up together. In the morning I can get the actual details. At just 3B tokens, Boris-2-0917, as I’ve taken to calling it, has already beaten the 30B checkpoint on all benchmarks.

replied to their post about 2 months ago

I know everybody was awaiting ASI but we gotta wait a couple more weeks I’m sorry

replied to Datdanboi25's post about 2 months ago

Wow! Incredible work!

reacted to Datdanboi25's post with 🔥 about 2 months ago

Post

3,667

THE SLM FRONTIER ADVANCES!

bench-labs/cagliostro-v3 just hit an Intelligence Index of 26.13 on the AxiomicLabs/Open_SLM_Leaderboard a 146M-param model trained completely from scratch on a single consumer GPU. That's 2nd place overall, and as far as I can tell, the most capable SLM trained on consumer hardware to date. Beating SmolLM-135m on 1/8th of the data is just silly levels of efficiency.

Big congrats to the @BenchLabs team and specifically @TobiasLogic !

6 replies

·

reacted to NILKNARFGonzo's post with 👀 about 2 months ago

Post

3,792

get played unsloth

gemma just deleted its own model runner with DeepSeek Harness

shoutout to deepseek and unsloth

11 replies

·

posted an update about 2 months ago

Post

269

Important Boris-2 news:

Boris-2 is 30B out of 200B tokens in, and it is severely behind its competitors in training.

We have determined the bug to be a configuration error. Boris-2 has been in training for ~1 week, and was projected to finish on November 3rd, 2026.

We are unfortunately going to restart training, with proper configuration.

The new projected finish date is ~15-18th of November.

We apologize for the delay.

17 replies

·

reacted to wayneworkman2012's post with 🔥 about 2 months ago

Post

108

Peacebell is released!

wayneworkman2012/peacebell-v1-291M
wayneworkman2012/peacebell-v1-291M-demo
wayneworkman2012/peacebell-v1-291M-demo-cpu
wayneworkman2012/peacebell-v1-148M

posted an update 2 months ago

Post

2,180

The SLM Consortium has begun work on a safety dataset for Small Language Models, with the creation of the dataset being headed by @wayneworkman2012

The dataset will focus on refusals and redirects surrounding dangerous or extreme sexual content, designed to be shaped sized appropriately for SLMs, without significantly lowering benchmark performance.

More info will be out soon!

3 replies

·

replied to Banaxi-Tech's post 2 months ago

It's great to see further adoption of n-gram! I think it really helps a lot.

reacted to Banaxi-Tech's post with 🔥 2 months ago

Post

3,046

We're going to release our BananaMind 2.1 models very soon!
We're also announcing 2 new models.

All of our models we will train are:
BananaMind 2.1 Flash Lite, 10M parameters with 8M in transformer and 2M in n-gram. 50B pretraining tokens.
BananaMind 2.1 Lite with 25M parameters, 5M in n-gram and 20M in transformer. 75B pretraining tokens.
BananaMind 2.1 Flash with 50M parameters, with undecided n-gram count yet. 100B pretraining tokens.
BananaMind 2.1 Pro with 145M parameters, with undecided n-gram count yet. 150-200B pretraining tokens.
BananaMind 2.1 Coder with 149M parameters with undecided n-gram count yet.
We're now announcing BananaMind 2.1 NanoCoder, a 10M parameter model focused specifically on coding and BananaMind 2.1 MiniCoder which is a 25M parameter model focused on coding.

Follow us:

BananaMind
@Banaxi-Tech
@vovaRL
@DedeProGames

bananamind-research-community

5 replies

·