8bpw 8h
Frostwind-v1
A finetune of upstage/SOLAR-10.7B-v1.0
Took Roughly 3 Hours with 4x 4090s, over 2 Epochs, with around 52K varied samples.
Dataset Composition:
20% - Coding
30% - Instruct
30% - Generalised Data
10% - Roleplay
10% - Dealignment
Testing Notes:
Fairly smart, as I expected. Obviously not at the level of the bigger models, but I did not expect that level from this.
Could be sampler issues, but generally I needed 1/2 swipes to get the correct answer when doing Zero context tests. If context is filled, no issues on my end.
For Roleplays: adding things like avoid writing as {{user}} suprisingly helps. Plus a proper prompt of course. I liked the writing style. Handles group characters in 1 card well, during my tests.
Fairly uncensored during roleplay. Yeah the as an AI stuff can happen at Zero context, but I have no issues once a character card is introduced. I had no issues making outputs that would give me 2500 Life Sentences if posted here.
Trained with Alpaca Format:
### Instruction:
<Prompt>
### Response:
OR
### Instruction:
<Prompt>
### Input:
<Insert Context Here>
### Response:
wandb:
wandb: Run history:
wandb: eval/loss ββββββββββββββββββββ
wandb: eval/runtime ββββββββββββββββββββ
wandb: eval/samples_per_second ββββββββββββββββββββ
wandb: eval/steps_per_second ββββββββββββββββββββ
wandb: train/epoch βββββββββββββββββββββ
β
β
β
β
β
ββββββββββββββ
wandb: train/global_step βββββββββββββββββββββ
β
β
β
β
β
ββββββββββββββ
wandb: train/learning_rate ββββββββββββββββββ
β
β
β
βββββββββββββββββββ
wandb: train/loss ββ
β
ββ
β
ββββββ
ββββ
βββ
β
β
βββββββββββββββββββ
wandb:
wandb: Run summary:
wandb: eval/loss 0.74622
wandb: eval/runtime 72.5049
wandb: eval/samples_per_second 37.239
wandb: eval/steps_per_second 2.331
wandb: train/epoch 1.98
wandb: train/global_step 410
wandb: train/learning_rate 0.0
wandb: train/loss 0.6457
wandb: train/total_flos 3.4382652340646707e+18
wandb: train/train_loss 0.70204
wandb: train/train_runtime 10880.917
wandb: train/train_samples_per_second 9.417
wandb: train/train_steps_per_second 0.038
wandb:
- Downloads last month
- 16