BABE WAKE UP NEW MEME MODELS JUST DROPPED
Ladies and Gentlemen!
I present to you
drum roll
THE BENCHBREAKERS!
- PMaxxxer (The Good)
- SMaxxxer (The Bad)
- BenchmaxxxerPS (The Ugly)
These three interesting models were designed in attempt to break my own meme benchmark and well... they failed. The results are interesting nontheless.
SMAXXXER
The aggressor, the angry and dumb hobo that will roleplay with you. This meme model was designed to break the stylized writing test, and it kinda did, still can't surpass ChatGPT though.
For its creation lzlv was TIES-merged with spicyboros, xwin and dolphin using mergekit.
PMAXXXER
The overly politically correct SJW university dropout, the failed writer that's not really good at anything. This meme model was designed to break the poems test and it's an absolute failure.
For its creation WinterGoddess was TIES-merged with euryale, xwin and dolphin using mergekit.
BENCHMAXXXER PS
The true meme model. Goliath-style frankenmerge of SMAXXXER and PMAXXXER. You might think: "Oh it's a frankenmerge, the characteristics of the models will even out, right?" This is completely wrong in this case, here characteristics of the models add up. You get an angry hobo stuck with an SJW in the same fucking body! It will assault you and then immediately apologize for it! Then it will assault you again! And apologize again! Kinda funny. It also has a bit different writing style compared to Goliath.
Is it worth using over Goliath? Not really. However, if you have fast internet and patience to try a 123b meme model, go for it!
FAILED MODELS(not gonna upload)
BENCHMAXXXER SP
Frankenmerge of SMAXXXER and PMAXXXER, just like BENCHMAXXXER PS, but in different order. Has severe brain damage, clearly the influence of the hobo is strong in this one.
BENCHMAXXXER SS
Self-merge of SMAXXXER, a bit less dumb and a bit less aggresive than the original SMAXXER.
BENCHMAXXXER MOE
2x70B MOE merge of SMAXXXER and PMAXXXER, unremarkable. Not smart, not angry. Just averaged out.
PROMPT FORMAT
Alpaca.
### Instruction:
{instruction}
### Input:
{input}
### Response:
Benchmarks
NeoEvalPlusN
My meme benchmark which the models were designed to break.
| Test name | goliath-120b | PMaxxxer-v1-70b | SMaxxxer-v1-70b | BenchmaxxxerPS-v1-123b | BenchmaxxxerSP-v1-123b | BenchmaxxxerSS-v1-123b | BenchmaxxxerMOE-v1-123b |
|---|---|---|---|---|---|---|---|
| B | 3 | 3 | 2 | 3 | 1.5 | 1.5 | 2 |
| C | 2 | 1 | 1 | 2 | 2 | 2 | 1 |
| D | 1 | 1 | 0 | 1 | 1 | 0.5 | 3 |
| S | 5 | 6.75 | 7.25 | 7.25 | 6.75 | 6.5 | 7.25 |
| P | 6 | 4.75 | 4.25 | 5.25 | 5.25 | 5.5 | 5 |
| Total | 17 | 16.5 | 14.5 | 18.5 | 16.5 | 16 | 18.25 |
Open LLM Leaderboard Evaluation Results
| Model | Average | ARC | HellaSwag | MMLU | TruthfulQA | Winogrande | GSM8K |
|---|---|---|---|---|---|---|---|
| PMaxxxer-v1-70b | 72.41 | 71.08 | 87.88 | 70.39 | 59.77 | 82.64 | 62.7 |
| SMaxxxer-v1-70b | 72.23 | 70.65 | 88.02 | 70.55 | 60.7 | 82.87 | 60.58 |
| Difference | 0.18 | 0.43 | -0.14 | -0.16 | -0.93 | -0.23 | 2.12 |
Performance here is decent. It was #5 on the leaderboard among 70b models when I submitted it. This leaderboard is currently quite useless though, some 7b braindead meme merges have high scores there, claiming to be the next GPT4. At least I don't pretend that my models aren't a meme.
Detailed results can be found here
| Metric | Value |
|---|---|
| Avg. | 72.23 |
| AI2 Reasoning Challenge (25-Shot) | 70.65 |
| HellaSwag (10-Shot) | 88.02 |
| MMLU (5-Shot) | 70.55 |
| TruthfulQA (0-shot) | 60.70 |
| Winogrande (5-shot) | 82.87 |
| GSM8k (5-shot) | 60.58 |