Spaetzle
Collection
German-English models, mostly merged, some sft/dpo
•
107 items
•
Updated
axolotl version: 0.4.0
base_model: cstr/Spaetzle-v8-7b
model_type: MistralForCausalLM
tokenizer_type: LlamaTokenizer
load_in_8bit: false
load_in_4bit: false
strict: false
rl: orpo
chat_template: chatml
datasets:
- path: cstr/ultrafeedback-binarized-preferences-cleaned-de-2
type: orpo.chat_template
chat_template: chatml
dataset_prepared_path:
val_set_size: 0.0
output_dir: ./out
remove_unused_columns: false
sequence_len: 1024
sample_packing: false
pad_to_sequence_len: false
eval_sample_packing: false
special_tokens:
eos_token: "<|im_end|>"
tokens: # these are delimiters
- "<|im_start|>"
gradient_accumulation_steps: 8
micro_batch_size: 1
num_epochs: 1
optimizer: paged_adamw_8bit
lr_scheduler: cosine
learning_rate: 5e-6
orpo_alpha: 0.1
save_safetensors: true
wandb_project: Spaetzle-v8-7b-orpo
wandb_entity:
wandb_watch:
wandb_name:
wandb_log_model:
train_on_inputs: false
group_by_length: false
bf16: auto
fp16:
tf32: false
gradient_checkpointing: true
gradient_checkpointing_kwargs:
use_reentrant: true
early_stopping_patience:
resume_from_checkpoint:
local_rank:
logging_steps: 1
xformers_attention:
flash_attention: true
warmup_steps: 100
evals_per_epoch:
eval_table_size:
eval_max_new_tokens: 128
saves_per_epoch: 4
debug:
deepspeed:
weight_decay: 0.0
fsdp:
fsdp_config:
This model is a fine-tuned version of cstr/Spaetzle-v8-7b on the None dataset.
More information needed
More information needed
More information needed
The following hyperparameters were used during training: