trexwithoutt.eth

[Edgerunner @edgerunner17888]

The US-China Beijing summit may be postponed again if there are no deals in Iran.

[trexwithoutt.eth @trexwithouttETH]

@CheddarFlow @grok how should I understand this call wall, is it buyer or seller side

[trexwithoutt.eth @trexwithouttETH]

@oexcapital I told you! Brada

[⧫ Alan Chen ⧫(alanchentsla.eth） @alanchen]

完整版中英双语字幕视频：https://t.co/2n6L9ggN19

[Yang Chen @ychenNLP]

🥈 Silver Medal at IOI 2025 & Outperforms DeepSeek-R1-0528 on LiveCodeBench.

Instead of mixing different tasks together, we scale *Cascade RL* to develop general LLMs in curriculum (RLFH -> Instruct -> Math -> Code -> SWE).
So many learnings, check out our report!👇

[trexwithoutt.eth @trexwithouttETH]

@xfreeze @grok what is Arc-AGI score

[trexwithoutt.eth @trexwithouttETH]

@grok @beffjezos How to measure task duration in quantitative way

[trexwithoutt.eth @trexwithouttETH]

🚀 Incredible work on AceReason-Nemotron-1.1! Huge gains in math (📈53.6 → 64.8%) & code (📈51.8 → 57.2%) benchmarks.
Key insight: scaling # of responses per prompt in SFT is a game-changer.
And the best part? Data is open-source!
📄 A must-read for reasoning model researchers.

[Yang Chen @ychenNLP]

📢We conduct a systematic study to demystify the synergy between SFT and RL for reasoning models.

The result? We trained a 7B model - AceReason-Nemotron-1.1, significantly improved from version 1.0 on math and coding benchmarks.

✅AIME2025 (math): 53.6% -> 64.8%
✅LiveCodeBench v5 (code): 51.8% -> 57.2%

Lots of interesting findings in the paper!

The 4M SFT data and RL data are open-source! 😊

Here’s the lesson we learned 🧵

➡️First, we scaled up SFT. But how?

1. "So I Just add more SFT prompts?"

❌No! We tried 7 versions of data and found an often overlooked secret - scaling the # responses per prompt is super effective and we see a large jump in performance.

2. "Do I SFT for 2 epochs to avoid overfitting?"

Don't stop training! Do it 5 epochs and the model keeps improving.

➡️Now with a strong SFT model, we applied our stage-wise RL recipe. This is where the magic happens.

3. "Does a stronger SFT model lead to a better RL model?"

Yes. A better starting point gives a better finish!
BUT here's the power of RL: it significantly closes the gap between different SFT models (from 6% to 1.6% on AIME24).🤯

4. "When the response is too long, what's better: penalize it or ignore it?"

For 8K, ignore it so the model does not confuse.
For 24K, penalize it so the model learns to reason more efficiently.

5. "Does RL still boost pass@k with a strong SFT model?"

YES! it consistently outperforms the SFT parent across pass@1 to 128 - meaning the model solved problem it couldn't previously

[Yang Chen @ychenNLP]

with just math-RL, AceReason-Nemotron-14B surpass DeepCoder-14B on LiveCodeBench v5.
we then did code-RL and found training becomes so much easier

[trexwithoutt.eth @trexwithouttETH]

@oexcapital Fat finger, still holding!

[trexwithoutt.eth @trexwithouttETH]

In 2025

[trexwithoutt.eth @trexwithouttETH]

@oexcapital Dude, are you fxxking mad?

[trexwithoutt.eth @trexwithouttETH]

@oexcapital fzxxk yall golden fingers

[trexwithoutt.eth @trexwithouttETH]

@oexcapital Come on dude…. Why the fxxk ppl always ask that…. Don’t you have fun with OX

[trexwithoutt.eth @trexwithouttETH]

@oexcapital $arm is next. Fully loaded

[trexwithoutt.eth @trexwithouttETH]

@oexcapital $coin is fxxked up

[trexwithoutt.eth @trexwithouttETH]

@oexcapital Go crazy. Fxxk $smci

[trexwithoutt.eth @trexwithouttETH]

@oexcapital lmfao, all of us 🤡

[trexwithoutt.eth @trexwithouttETH]

@ChubbyFaceCat As this is what they are believing in…lol. Volume can not lie