New model palmer-006
palmer-006
Benchmark Accuracy
ARC Challenge 29.27%
ARC Easy 44.82%
PIQA 63.60%
Winogrande 50.36%
HellaSwag 38.41%
Arithmark-3 52.70%
You need make a PR
You need make a PR
Yes pls, pr with acc norm results from lm eval harness
You need make a PR
Yes pls, pr with acc norm results from lm eval harness
On it
Done
Hi appvoid, out of curiosity how was the merged final model created?
The final model was selected because the SPINOFF (a new technique I'm developing) and KD directions produced the strongest verified combined result without further training. It became a better language model at the expense of "un-instructing". The goal was basically to become good at bananamind without leaking examples by training it on totally unrelated tasks until I found one that generalizes to everything.
So I found that combining hard mining DPO negatives and combining to a more general language modeling branch provides best results. I'll try to make a blog on these findings whenever I have time.
Check this overview I feeded all my files to GPT-5:
tiiuae/Falcon-H1-Tiny-90M-Instruct
│
├── SPINOFF branch
│ ├── 192 unrelated open-ended preference prompts
│ ├── Strong public answers used as chosen
│ ├── Student-generated answers used as rejected
│ ├── LoRA preference training
│ └── SPIN v1: 229/350
│
├── Knowledge-distillation branch
│ ├── Falcon-H1 0.5B Instruct teacher
│ ├── Decontaminated prose/code corpus
│ ├── 48,768 cached tokens
│ ├── Top-32 teacher probabilities
│ ├── LoRA r=8, LR 3e-6
│ ├── Non-recurrent attention/MLP adaptation
│ └── KD v2: 230/350
│
└── 50:50 parameter average
└── Palmer-006: 231/350, 66.00%
hmm ok should be fine then, cool stuff!