See axolotl config

axolotl version: 0.4.0

base_model: andysalerno/mistral-sft-v3
model_type: AutoModelForCausalLM

load_in_8bit: true
load_in_4bit: false
strict: false

datasets:
  - path: andysalerno/rainbowfish-v1
    type:
      system_prompt: ""
      field_system: system
      field_instruction: input
      field_output: output
      format: "{instruction}"
      no_input_format: "{instruction}"
dataset_prepared_path: last_run_prepared
val_set_size: 0.005
output_dir: ./lora-out-rainbow9

adapter: lora
lora_model_dir:

sequence_len: 2048
sample_packing: false # was true
eval_sample_packing: false
pad_to_sequence_len: false
padding_side: left

lora_r: 64
lora_alpha: 16
lora_dropout: 0.05
lora_target_linear: true
lora_fan_in_fan_out:
lora_target_modules:
  - gate_proj
  - down_proj
  - up_proj
  - q_proj
  - v_proj
  - k_proj
  - o_proj

lora_modules_to_save:
  - embed_tokens
  - lm_head

wandb_project: axolotl
wandb_entity:
wandb_watch:
wandb_name:
wandb_log_model:

gradient_accumulation_steps: 4
micro_batch_size: 4
optimizer: paged_adamw_8bit
lr_scheduler: cosine
learning_rate: 2e-5

neftune_noise_alpha: 5

train_on_inputs: false
group_by_length: false
bf16: true
fp16:
tf32: false

gradient_checkpointing: true
gradient_checkpointing_kwargs:
  use_reentrant: false
# early_stopping_patience: 3
local_rank:
logging_steps: 1
xformers_attention:
flash_attention: true

loss_watchdog_threshold: 5.0
loss_watchdog_patience: 3

hub_strategy: "every_save"
hub_model_id: andysalerno/rainbowfish-v9-adapter

num_epochs: 4
warmup_steps: 100 
eval_steps: 200
eval_table_size:
eval_table_max_new_tokens: 128
# max_steps: 500
saves_per_epoch: 1
debug:
weight_decay: 0.1
fsdp:
fsdp_config:
special_tokens:
  bos_token: "<|im_start|>"
  eos_token: "<|im_end|>"
  unk_token: "<unk>"

rainbowfish-v9-adapter

This model is a fine-tuned version of andysalerno/mistral-sft-v3 on the None dataset. It achieves the following results on the evaluation set:

• Loss: 0.6456

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: 2e-05
• train_batch_size: 4
• eval_batch_size: 4
• seed: 42
• distributed_type: multi-GPU
• num_devices: 4
• gradient_accumulation_steps: 4
• total_train_batch_size: 64
• total_eval_batch_size: 16
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: cosine
• lr_scheduler_warmup_steps: 100
• num_epochs: 4

Training results

Training Loss	Epoch	Step	Validation Loss
0.6535	0.18	200	0.6840
0.69	0.37	400	0.6711
0.6649	0.55	600	0.6641
0.6959	0.74	800	0.6590
0.717	0.92	1000	0.6547
0.5243	1.11	1200	0.6540
0.6285	1.29	1400	0.6523
0.6219	1.47	1600	0.6504
0.6334	1.66	1800	0.6486
0.6627	1.84	2000	0.6466
0.6319	2.03	2200	0.6460
0.6081	2.21	2400	0.6466
0.5721	2.4	2600	0.6459
0.5794	2.58	2800	0.6447
0.721	2.76	3000	0.6443
0.5825	2.95	3200	0.6436
0.5921	3.13	3400	0.6457
0.5224	3.32	3600	0.6461
0.5466	3.5	3800	0.6456
0.5972	3.69	4000	0.6460
0.5999	3.87	4200	0.6456

Framework versions

• PEFT 0.8.2
• Transformers 4.38.0.dev0
• Pytorch 2.1.2+cu118
• Datasets 2.17.0
• Tokenizers 0.15.0