See axolotl config

axolotl version: 0.4.0

base_model: premai-io/prem-1B-chat
model_type: LlamaForCausalLM
tokenizer_type: AutoTokenizer

load_in_8bit: false
load_in_4bit: false
strict: false

datasets:
  - path: argilla/distilabel-capybara-dpo-7k-binarized
    type: orpo.chat_template
dataset_prepared_path: last_run_prepared
val_set_size: 0.001
output_dir: ./prem-1B-chat-32k
save_safetensors: true
sequence_len: 8192
sample_packing: false
pad_to_sequence_len: false
use_pose: true
pose_max_context_len: 262144
min_sample_len: 6144
pose_num_chunks: 16
curriculum_sampling: true

overrides_of_model_config:
  rope_theta: 500000.0
  max_position_embeddings: 262144

  # peft_use_dora: true
adapter: lora
peft_use_rslora: true
lora_model_dir:
lora_r: 1024
lora_alpha: 1024
lora_dropout: 0.1
lora_target_modules:
  - q_proj
  - k_proj
  - v_proj
  - o_proj
lora_modules_to_save:
  - embed_tokens
  - lm_head

wandb_project:
wandb_entity: 
wandb_watch:
wandb_name:
wandb_log_model:

gradient_accumulation_steps: 8
micro_batch_size: 1
num_epochs: 20
optimizer: adamw_bnb_8bit
lr_scheduler: cosine
learning_rate: 0.00001
max_grad_norm: 1.0
adam_beta2: 0.95

train_on_inputs: false
group_by_length: false
bf16: true
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
sdp_attention:
s2_attention:

warmup_steps: 10
evals_per_epoch: 8
saves_per_epoch: 8
debug:
deepspeed:
weight_decay: 0.0
fsdp:
fsdp_config:
special_tokens:
  pad_token: <|end_of_text|>

prem-1B-chat-32k

This model is a fine-tuned version of premai-io/prem-1B-chat on the Argilla DPO dataset. 32K context extended using PoSE method It achieves the following results on the evaluation set:

• Loss: 6.9843

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: 1e-05
• train_batch_size: 1
• eval_batch_size: 1
• seed: 42
• distributed_type: multi-GPU
• num_devices: 8
• gradient_accumulation_steps: 8
• total_train_batch_size: 64
• total_eval_batch_size: 8
• optimizer: Adam with betas=(0.9,0.95) and epsilon=1e-08
• lr_scheduler_type: cosine
• lr_scheduler_warmup_steps: 10
• num_epochs: 20

Training results

Training Loss	Epoch	Step	Validation Loss
0.5389	1.0	1	6.3469
0.5389	2.0	2	6.2533
0.5017	3.0	3	6.2101
0.4689	4.0	4	6.3163
0.3604	5.0	5	6.4144
0.3107	6.0	6	6.4127
0.2698	7.0	7	6.8089
0.317	8.0	8	7.3388
0.2228	9.0	9	6.5063
0.1798	10.0	10	5.7073
0.1436	11.0	11	5.1185
0.1183	12.0	12	4.8994
0.1002	13.0	13	4.8033
0.0865	14.0	14	5.1707
0.0758	15.0	15	5.7089
0.0663	16.0	16	6.4052
0.0601	17.0	17	6.7814
0.0545	18.0	18	6.9586
0.0505	19.0	19	6.9766
0.0482	20.0	20	6.9843

Framework versions

• PEFT 0.10.0
• Transformers 4.40.0.dev0
• Pytorch 2.1.2+cu121
• Datasets 2.15.0
• Tokenizers 0.15.0