62 lines
1.5 KiB
Markdown
62 lines
1.5 KiB
Markdown
tldr; This is Phi 3 Medium finetuned for (mainly SFW) roleplaying.
|
|
|
|
It was a promising release candidate that fell flat when things got moist.
|
|
|
|
I'm publishing all the details for anyone else interested in finetuning Phi 3.
|
|
|
|
Training Details:
|
|
- 8x H100 80GB SXM GPUs
|
|
- 1 hour training time
|
|
|
|
Results for Roleplay Mode (i.e., not Instruct format):
|
|
- Strong RP formatting.
|
|
- Tends to output short, straightforward replies to the player character.
|
|
- Starts to break down when things get moist.
|
|
- Important: My testing is lazy and flawed. Take it with a grain of salt and test the GGUFs before taking notes.
|
|
|
|
|
|
|
|
Axolotl Config (some fields omitted)
|
|
```yaml
|
|
base_model: failspy/Phi-3-medium-4k-instruct-abliterated-v3
|
|
load_in_4bit: true
|
|
bf16: auto
|
|
fp16:
|
|
tf32: false
|
|
flash_attention: true
|
|
|
|
sequence_len: 4096
|
|
datasets:
|
|
- path: Undi95/andrijdavid_roleplay-conversation-sharegpt
|
|
type: customphi3
|
|
|
|
num_epochs: 2
|
|
warmup_steps: 30
|
|
weight_decay: 0.1
|
|
|
|
adapter: lora
|
|
lora_r: 128
|
|
lora_alpha: 16
|
|
lora_dropout: 0.1
|
|
lora_target_linear: true
|
|
|
|
gradient_accumulation_steps: 2
|
|
micro_batch_size: 2
|
|
gradient_checkpointing: true
|
|
gradient_checkpointing_kwargs:
|
|
use_reentrant: true
|
|
|
|
sample_packing: true
|
|
pad_to_sequence_len: true
|
|
|
|
optimizer: paged_adamw_8bit
|
|
lr_scheduler: cosine
|
|
learning_rate: 0.0001
|
|
max_grad_norm: 1.0
|
|
|
|
val_set_size: 0.01
|
|
evals_per_epoch: 3
|
|
eval_max_new_tokens: 128
|
|
eval_batch_size: 1
|
|
```
|