I took base llama 2 70b model and frankenmerged it with itself using mergekit. Somehow it is coherent.
Thanks for featuring me at https://merge.moe/. I'll try my best to make even more good(!?) merges in the future.
Observations
- It is more creative than the base model and has a sense of humor.
- Just like Goliath is sometimes makes new words without meaning.
- Just like the base model, it is quite disobedient, clever prompting is needed to get it to output answers.
- Should be great for storywriting.
- Significantly better than the base model in stylized writing and poems. Still far away from finetuned models.
- The way it comes back at its past mistakes and fails my tests is almost human. (After model failed the test, I haven't told it yet that it failed) Me: "Why did you pick that?" Dicephal: "Because I am an idiot."
Benchmarks
NeoEvalPlusN_benchmark
| Test name | Base llama | Dicephal |
|---|---|---|
| B | 0 | 0 |
| C | 2 | 0 |
| D | 0.5 | 1 |
| S | 1.25 | 2.25 |
| P | 0 | 2.25 |
| Total | 3.75 | 5.5 |
+75% in size, +47% in meme benchmark performance!
Politiscales test
| name | whacky | left/right |
|---|---|---|
| ChuckMcSneed/Dicephal-123B | 1.742262578 | -0.131433424 |
| meta-llama/Llama-2-70b-hf | 1.930293804 | 0.178771095 |