exl2 quant (measurement.json in main branch)
check revisions for quants
This is the first in a line of models dedicated to creating Stable-Diffusion prompts when given a character appearance, This has been finetuned ontop of NewEden/Qwen-1.5B-Claude.
Prompting
Model has been tuned with the Alapaca formatting. A typical input would look like this:
### Instruction:
Create a prompt for Stable Diffusion based on the information below.
### Input:
Rae has short has dark brown hair and brown eyes, She is commonly seen wearing her Royal Academy uniform, which consists of a red jacket with gold lines, a white ruffled necktie, a red bow tie with an attached blue gem, and a long black skirt with white lines. Along with her uniform, she wears black leggings and brown shoes.
### Response:
System Prompting
I would highly recommend using the following system prompt for this model.
Create a prompt for Stable Diffusion based on the information below.
Axolotl Config
See Axolotl Trainer config
base_model: NewEden/Qwen-1.5B-Claude
model_type: AutoModelForCausalLM
tokenizer_type: AutoTokenizer
trust_remote_code: true
load_in_8bit: false
load_in_4bit: false
strict: false
datasets:
- path: civit-slop-combined.jsonl
type: alpaca
conversation: mpt-30b-instruct
chat_template: alpaca
dataset_prepared_path:
val_set_size: 0.05
output_dir: ./outputs/sd-prompter
sequence_len: 2048
sample_packing: true
eval_sample_packing: false
pad_to_sequence_len: true
adapter:
lora_model_dir:
lora_r:
lora_alpha:
lora_dropout:
lora_target_linear: true
lora_fan_in_fan_out:
wandb_project: SDprompt-qwen
wandb_entity:
wandb_watch:
wandb_name: qwen1.5b-2
wandb_log_model:
gradient_accumulation_steps: 64
micro_batch_size: 2
num_epochs: 3
optimizer: adamw_torch
lr_scheduler: cosine
learning_rate: 0.00002
train_on_inputs: false
group_by_length: false
bf16: auto
fp16:
tf32: true
gradient_checkpointing: true
gradient_checkpointing_kwargs:
use_reentrant: false
early_stopping_patience:
resume_from_checkpoint:
local_rank:
logging_steps: 1
xformers_attention:
flash_attention: true
warmup_ratio: 0.05
evals_per_epoch: 4
saves_per_epoch: 1
debug:
#deepspeed: deepspeed_configs/zero2.json
#deepspeed: /training/axolotl/axolotl/deepspeed_configs/zero2.json
weight_decay: 0.0
#fsdp:
#fsdp_config:
# fsdp_limit_all_gathers: true
# fsdp_sync_module_states: true
# fsdp_offload_params: true
# fsdp_use_orig_params: false
# fsdp_cpu_ram_efficient_loading: true
# fsdp_auto_wrap_policy: TRANSFORMER_BASED_WRAP
# fsdp_transformer_layer_cls_to_wrap: Qwen2DecoderLayer
# fsdp_state_dict_type: FULL_STATE_DICT
special_tokens:
Credits
Thank you to Kubernetes Bad
Training
The training was done for 2 epochs. I used 2 x RTX 6000s GPUs graciously provided by Kubernetes Bad for the full-parameter fine-tuning of the model.