---
license: apache-2.0
library_name: peft
tags:
- axolotl
- generated_from_trainer
base_model: byroneverson/LLaVA-v1.5-7B-rehome
model-index:
- name: LLaVA-v1.5-7B-rehome-qlora
  results: []
---

<!-- This model card has been generated automatically according to the information the Trainer had access to. You
should probably proofread and complete it, then remove this comment. -->

[<img src="https://raw.githubusercontent.com/OpenAccess-AI-Collective/axolotl/main/image/axolotl-badge-web.png" alt="Built with Axolotl" width="200" height="32"/>](https://github.com/OpenAccess-AI-Collective/axolotl)
<details><summary>See axolotl config</summary>

axolotl version: `0.3.0`
```yaml
base_model: byroneverson/LLaVA-v1.5-7B-rehome
model_type: AutoModelForCausalLM #MistralForCausalLM
tokenizer_type: AutoTokenizer #LlamaTokenizer
#is_mistral_derived_model: true
#
load_in_8bit: false
load_in_4bit: true
strict: false
#
datasets:
  - path: yahma/alpaca-cleaned #byroneverson/shell-cmd-instruct
    type: completion #solar_shell_instruct #alpaca
    # For `completion` datsets only, uses the provided field instead of `text` column
    field: output
dataset_prepared_path: last_run_prepared
val_set_size: 0.02
output_dir: /kaggle/working/qlora-out
#
# Push checkpoints to hub
hub_model_id: byroneverson/LLaVA-v1.5-7B-rehome-shell-qlora
# How to push checkpoints to hub
# https://huggingface.co/docs/transformers/v4.31.0/en/main_classes/trainer#transformers.TrainingArguments.hub_strategy
hub_strategy: checkpoint
# Whether to use hf `use_auth_token` for loading datasets. Useful for fetching private datasets
# Required to be true when used in combination with `push_dataset_to_hub`
hf_use_auth_token: true
#
adapter: qlora
lora_model_dir:
#
sequence_len: 2048
sample_packing: true
eval_sample_packing: false
pad_to_sequence_len: true
# Train only these params when performing full train
#unfrozen_parameters: [
#    "\\bmlp\\.up_proj\\b",
#    "\\bmlp\\.down_proj\\b",
#    "\\bmlp\\.gate_proj\\b",
#  ]
# LoRA/QLoRA
lora_r: 160 #128
lora_alpha: 80 #384 # A good ballpark is double lora_r for 2x or half for 0.5x
lora_dropout: 0.025
lora_target_linear: false # Only target specified modules
lora_fan_in_fan_out:
lora_target_modules: [
    "up_proj",
    "down_proj",
    "gate_proj",
  ]
#
wandb_project: "LLaVA-v1.5-7B-rehome-qlora"
wandb_log_model: "checkpoint"
wandb_entity:
wandb_watch:
wandb_run_id:
#
gradient_accumulation_steps: 16 # 1
micro_batch_size: 1
num_epochs: 0.2
optimizer: paged_lion_8bit #paged_adamw_8bit
lr_scheduler: cosine
cosine_min_lr_ratio: 1.0 # Constant lr if 1.0
learning_rate: 0.0001
#
train_on_inputs: false
group_by_length: false
bf16: false #true
fp16: true
tf32: false
#
gradient_checkpointing: true
early_stopping_patience:
# Resume from a specific checkpoint dir
resume_from_checkpoint: #last-checkpoint
# If resume_from_checkpoint isn't set and you simply want it to start where it left off.
# Be careful with this being turned on between different models.
auto_resume_from_checkpoints: true #false
local_rank:
logging_steps: 1
xformers_attention:
# Whether to use flash attention patch https://github.com/Dao-AILab/flash-attention:
flash_attention: false #true
flash_attn_cross_entropy:  # Whether to use flash-attention cross entropy implementation - advanced use only
flash_attn_rms_norm: false # Whether to use flash-attention rms norm implementation - advanced use only
flash_attn_fuse_qkv: # Whether to fuse QKV into a single operation
flash_attn_fuse_mlp: # Whether to fuse part of the MLP into a single operation
#
warmup_steps: 4
eval_steps: 50
eval_table_size:
eval_table_max_new_tokens: 128
save_steps:
debug: true #
deepspeed:
weight_decay: 0.0
fsdp:
fsdp_config:
special_tokens:
  bos_token: "<s>"
  eos_token: "</s>"
  unk_token: "<unk>"

```

</details><br>

# LLaVA-v1.5-7B-rehome-shell-qlora

This model is a fine-tuned version of [byroneverson/LLaVA-v1.5-7B-rehome](https://huggingface.co/byroneverson/LLaVA-v1.5-7B-rehome) on the None dataset.

## 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
- distributed_type: multi-GPU
- num_devices: 2
- gradient_accumulation_steps: 16
- total_train_batch_size: 32
- total_eval_batch_size: 2
- optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
- lr_scheduler_type: cosine
- lr_scheduler_warmup_steps: 4
- num_epochs: 0.2
- mixed_precision_training: Native AMP

### Training results



### Framework versions

- PEFT 0.7.2.dev0
- Transformers 4.37.0.dev0
- Pytorch 2.0.1+cu117
- Datasets 2.16.1
- Tokenizers 0.15.0