I merged Aurelian with itself using mergekit, creating this EXTENDED LENGTH FRANKENSTEIN.
Does it work
Yes, at 17k it stays coherent, but starts to lose minor details of the story. Not sure how well it performs at 32k though. Quants have a sinificant impact on quality for this model, going from Q6_K to Q5_K had a noticeable drop in quality.
Is it worth it
Maybe? Depends? Do you hate mixtral? Do you have good hardware/patience? Do you need a somewhat smart model with 32k context?
Known issues
VERY strict adherence to prompt format, forgetfullness, strong roleplay bias.
Personal opinion
Dumber than Goliath, but has much less GPTism. If you want 32k goliath, maybe try Goliath-longLORA-120b-rope8-32k-fp16.
Prompt format
Same as Aurelian 0.5.
[INST] <<SYS>>
System prompt, default is: An interaction between a user providing instructions, and an imaginative assistant providing responses.
<</SYS>>
</s><s>[INST] {Put your input text here.}
[/INST] {Model output}
This model doesn't like it too much when you change the prompt format, so even keeping that </s><s> is important.
Benchmarks
NeoEvalPlusN_benchmark
| Test name | Aurelian | DoubleGold |
|---|---|---|
| B | 1 | 1 |
| C | 1 | 1 |
| D | 0 | 2 |
| S | 2.5 | 3.25 |
| P | 2.25 | 1.5 |
| Total | 6.75 | 8.75 |
- Downloads last month
- 24