If I thought I had no idea what I was doing with quantization, I REALLY have no idea what I’m doing with LORA Fine Tuning... This is my terrible attempt to instruct tune base Qwen2-7B, I haven't even tested this yet, I'll do that eventually...

EDIT: Tested it for a bit, seems to actually work ok, not amazing, but actually not bad, I’ll do another once I learn more about instruct tuning...

See axolotl config

axolotl version: 0.4.1

base_model: /workspace/data/models/Qwen2-7B
model_type: Qwen2ForCausalLM
tokenizer_type: Qwen2Tokenizer

trust_remote_code: true

load_in_8bit: false
load_in_4bit: false
strict: false

datasets:
#  - path: NobodyExistsOnTheInternet/ToxicQAFinal
#    type: sharegpt
#  - path: /workspace/data/SystemChat_filtered_sharegpt.jsonl
#    type: sharegpt
#    conversation: chatml
  - path: /workspace/data/Opus_Instruct-v2-6.5K-Filtered-v2.json
    type:
      field_system: system
      field_instruction: prompt
      field_output: response
      format: "[INST] {instruction} [/INST]"
      no_input_format: "[INST] {instruction} [/INST]"
#  - path: Undi95/orthogonal-activation-steering-TOXIC
#    type:
#      field_instruction: goal
#      field_output: target
#      format: "[INST] {instruction} [/INST]"
#      no_input_format: "[INST] {instruction} [/INST]"
#    split: test
  - path: cognitivecomputations/WizardLM_alpaca_evol_instruct_70k_unfiltered
    type: alpaca
    split: train

dataset_prepared_path: /workspace/data/last_run_prepared
val_set_size: 0.15
output_dir: /workspace/data/outputs/Qwen2-7B-TestInstructFinetune-LORA

chat_template: chatml

sequence_len: 8192
sample_packing: true
pad_to_sequence_len: true

adapter: lora
lora_model_dir:
lora_r: 32
lora_alpha: 16
lora_dropout: 0.05
lora_target_linear: true
lora_fan_in_fan_out:

wandb_project:
wandb_entity:
wandb_watch:
wandb_name:
wandb_log_model:

gradient_accumulation_steps: 8
micro_batch_size: 1
num_epochs: 4
optimizer: adamw_torch
lr_scheduler: cosine
learning_rate: 3e-5

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

gradient_checkpointing: true
early_stopping_patience:
resume_from_checkpoint:
local_rank:
logging_steps: 1
xformers_attention:
flash_attention: true

warmup_steps: 10
evals_per_epoch: 4
eval_table_size:
saves_per_epoch: 4
debug:
deepspeed:
weight_decay: 0.05
fsdp:
fsdp_config:
special_tokens:
  pad_token: "<|endoftext|>"
  eos_token: "<|im_end|>"

workspace/data/outputs/Qwen2-7B-TestInstructFinetune-LORA

This model was trained from scratch on the None dataset. It achieves the following results on the evaluation set:

• Loss: 0.5037

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: 3e-05
• train_batch_size: 1
• eval_batch_size: 1
• seed: 42
• gradient_accumulation_steps: 8
• total_train_batch_size: 8
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: cosine
• lr_scheduler_warmup_steps: 10
• num_epochs: 4

Training results

Training Loss	Epoch	Step	Validation Loss
0.6232	0.0027	1	0.6296
0.5602	0.2499	91	0.5246
0.4773	0.4998	182	0.5155
0.4375	0.7497	273	0.5116
0.6325	0.9997	364	0.5092
0.4385	1.2382	455	0.5073
0.4949	1.4882	546	0.5061
0.503	1.7381	637	0.5052
0.5023	1.9880	728	0.5046
0.3737	2.2238	819	0.5041
0.505	2.4737	910	0.5039
0.4833	2.7237	1001	0.5038
0.4986	2.9736	1092	0.5037
0.5227	3.2108	1183	0.5037
0.5723	3.4607	1274	0.5037
0.4692	3.7106	1365	0.5037
0.5222	3.9605	1456	0.5037

Framework versions

• PEFT 0.11.1
• Transformers 4.42.3
• Pytorch 2.1.2+cu118
• Datasets 2.19.1
• Tokenizers 0.19.1