See axolotl config

axolotl version: 0.4.1

base_model: lordspline/mergestein
model_type: AutoModelForCausalLM
tokenizer_type: AutoTokenizer

load_in_8bit: false
load_in_4bit: false
strict: false

chat_template: chatml
datasets:
  - path: lordspline/scidata
    type: sharegpt
    conversation: chatml
  - path: lordspline/wizard_v2_196k_unfiltered
    type: sharegpt
    conversation: chatml
  - path: lordspline/ultrainteract
    type: sharegpt
    conversation: chatml
  
dataset_prepared_path: last_run_prepared
val_set_size: 0.002

output_dir: ./mergestein

sequence_len: 8192
sample_packing: true
pad_to_sequence_len: true
eval_sample_packing: false

wandb_project: mergestein
wandb_entity:
wandb_watch:
wandb_name:
wandb_log_model:
hub_model_id: lordspline/mergestein

gradient_accumulation_steps: 1
micro_batch_size: 1
num_epochs: 1
optimizer: paged_adamw_8bit
lr_scheduler: cosine
learning_rate: 0.0001 # look

train_on_inputs: false
group_by_length: false
bf16: auto
fp16:
tf32: false

gradient_checkpointing: unsloth
gradient_checkpointing_kwargs:
   use_reentrant: true # look
early_stopping_patience:
resume_from_checkpoint: # ./mergestein/checkpoint-8015
local_rank:
logging_steps: 1
xformers_attention:
flash_attention: true

warmup_steps: 10
evals_per_epoch: 1
eval_table_size:
eval_max_new_tokens: 128
saves_per_epoch: 300
debug:

# deepspeed: deepspeed_configs/zero3_bf16.json
weight_decay: 0.05
fsdp:
fsdp_config:
special_tokens:
  eos_token: "<|im_end|>"
  pad_token: "<|end_of_text|>"
tokens:
  - "<|im_start|>"
  - "<|im_end|>"

mergestein

This model is a fine-tuned version of lordspline/mergestein on the None dataset. It achieves the following results on the evaluation set:

• Loss: 1.2069

Model description

More information needed

Intended uses & limitations

More information needed

Training and evaluation data

More information needed

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 0.0001
• train_batch_size: 1
• eval_batch_size: 1
• seed: 42
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: cosine
• lr_scheduler_warmup_steps: 10
• num_epochs: 1

Training results

Training Loss	Epoch	Step	Validation Loss
1.4476	1.0	48435	1.2069

Framework versions

• Transformers 4.41.1
• Pytorch 2.1.0+cu118
• Datasets 2.19.1
• Tokenizers 0.19.1