qwen-21b-creative-qlora

<!-- 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/axolotl-ai-cloud/axolotl/main/image/axolotl-badge-web.png" alt="Built with Axolotl" width="200" height="32"/>

<details><summary>See axolotl config</summary>

axolotl version: 0.8.0.dev0

# === Start-up Commands ===
# curl -LsSf https://astral.sh/uv/install.sh | sh
# export PATH="$HOME/.local/bin:$PATH"
# git clone https://github.com/axolotl-ai-cloud/axolotl
# cd axolotl
# git checkout d8b4027200de0fe60f4ae0a71272c1a8cb2888f7
# uv venv
# source .venv/bin/activate
# uv pip install packaging ninja setuptools huggingface_hub[cli,hf_transfer]
# uv pip install "cut-cross-entropy[transformers] @ git+https://github.com/apple/ml-cross-entropy.git"
# uv pip install apollo-torch
# uv pip install --no-build-isolation -e .[flash-attn,deepspeed]
# uv pip install git+https://github.com/huggingface/transformers.git
# export HF_HUB_ENABLE_HF_TRANSFER=1
# huggingface-cli login --token $hf_key && wandb login $wandb_key
# axolotl preprocess qwen21-creative.yml
# axolotl train qwen21-creative.yml
# 39-43
# 36-40
# 25-29
# = 25-29 and 36-43

# curl -LsSf https://astral.sh/uv/install.sh | sh && export PATH="$HOME/.local/bin:$PATH" && git clone https://github.com/axolotl-ai-cloud/axolotl && cd axolotl && uv venv && source .venv/bin/activate && uv pip install torch==2.5.1 packaging ninja setuptools huggingface_hub[cli,hf_transfer] && uv pip install apollo-torch && uv pip install --no-build-isolation -e .[flash-attn,deepspeed] && uv pip install git+https://github.com/huggingface/transformers.git && export HF_HUB_ENABLE_HF_TRANSFER=1 && cd .. && huggingface-cli login --token $hf_key && wandb login $wandb_key

# === Model Configuration ===
base_model: Columbidae/Qwen2.5-21B-Experimental-E2
load_in_8bit: false
load_in_4bit: true

# === HF Configuration === 
hub_model_id: ToastyPigeon/qwen-21b-creative-qlora
hub_strategy: "every_save"

# === Training Setup ===
num_epochs: 1
micro_batch_size: 1
gradient_accumulation_steps: 4
sequence_len: 4096
sample_packing: true
pad_to_sequence_len: true

# === Evaluation ===
val_set_size: 100
evals_per_epoch: 10
eval_table_size:
eval_max_new_tokens: 256
eval_sample_packing: true

# === LoRA Configuration ===
adapter: qlora
lora_model_dir:
lora_r: 32
lora_alpha: 64
lora_dropout: 0.5
lora_target_linear: 
lora_fan_in_fan_out:
lora_target_modules:
  - gate_proj
  - down_proj
  - up_proj
  - q_proj
  - v_proj
  - k_proj
  - o_proj

#lora_mlp_kernel: true
#lora_qkv_kernel: true
#lora_o_kernel: true

# === Hyperparameter Configuration ===
#optimizer: apollo_adamw
optimizer: paged_ademamix_8bit
# Apollo-mini configuration:
#optim_args: "proj=random,rank=1,scale=128.0,scale_type=tensor,update_proj_gap=200"
# Regular Apollo configuration:
# optim_args: 
#optim_target_modules: all_linear
learning_rate: 1e-5
lr_scheduler: cosine
weight_decay: 0.01
warmup_ratio: 0.05


# === Data Configuration ===
datasets:
  - path: ToastyPigeon/mixed-instruct
    type: chat_template
    split: train[:300]
    field_messages: conversations
    message_field_role: from
    message_field_content: value
  - path: ToastyPigeon/some-rp-v2-4k
    split: train[:1000]
    type: chat_template
    field_messages: conversations
    message_field_role: from
    message_field_content: value
  - path: ToastyPigeon/gutenberg-sft
    split: train[500:1000]
    type: chat_template
    field_messages: conversations
    message_field_role: from
    message_field_content: value
  - path: ToastyPigeon/fujin-filtered-instruct
    split: train[:500]
    type: chat_template
    field_messages: conversations
    message_field_role: from
    message_field_content: value
  - path: ToastyPigeon/ali-books
    type: completion
    field: text
  - path: ToastyPigeon/disco-chat
    type: completion
    field: text
  - path: ToastyPigeon/adventure-combined-no-slop-matches
    split: train[:300]
    type: completion
    field: text

dataset_prepared_path: last_run_prepared
chat_template: tokenizer_default
# Example custom template:
# chat_template: jinja
# chat_template_jinja: |
#   {{- bos_token }}{%- for message in messages %}
#   {%- if message['role'] == 'system' %}
#   {{- '[SYSTEM_PROMPT]' + message['content'] + '[/SYSTEM_PROMPT]' }}
#   {%- elif message['role'] == 'user' %}
#   {{- '[INST]' + message['content'] + '[/INST]' }}
#   {%- elif message['role'] == 'assistant' %}
#   {{- message['content'] + eos_token }}
#   {%- endif %}
#   {%- endfor %}

# === Plugins ===
plugins:
  - axolotl.integrations.liger.LigerPlugin
#  - axolotl.integrations.cut_cross_entropy.CutCrossEntropyPlugin

# === Hardware Optimization ===
gradient_checkpointing: unsloth
gradient_checkpointing_kwargs:
  use_reentrant: false
liger_rope: true
liger_rms_norm: true
liger_glu_activation: true
liger_fused_linear_cross_entropy: true
unsloth_cross_entropy_loss: true
#cut_cross_entropy: true
# Only if using multiple GPUs:
#deepspeed: axolotl/deepspeed_configs/zero3_bf16.json

# === Wandb Tracking ===
wandb_project: Qwen-27
# wandb_entity: [WANDB_ENTITY]
# wandb_name: [WANDB_RUN_NAME]

# === Checkpointing ===
saves_per_epoch: 10
save_total_limit: 1

# === Advanced Settings ===
output_dir: ./ckpts
bf16: auto
flash_attention: true
train_on_inputs: false
group_by_length: false
save_safetensors: true
logging_steps: 1
gc_steps: 10
seed: 69

</details><br>

qwen-21b-creative-qlora

This model is a fine-tuned version of Columbidae/Qwen2.5-21B-Experimental-E2 on the ToastyPigeon/mixed-instruct, the ToastyPigeon/some-rp-v2-4k, the ToastyPigeon/gutenberg-sft, the ToastyPigeon/fujin-filtered-instruct, the ToastyPigeon/ali-books, the ToastyPigeon/disco-chat and the ToastyPigeon/adventure-combined-no-slop-matches datasets. It achieves the following results on the evaluation set:

• Loss: 2.2767

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: 69
• distributed_type: multi-GPU
• num_devices: 4
• gradient_accumulation_steps: 4
• total_train_batch_size: 16
• total_eval_batch_size: 4
• optimizer: Use OptimizerNames.PAGED_ADEMAMIX_8BIT and the args are: No additional optimizer arguments
• lr_scheduler_type: cosine
• lr_scheduler_warmup_steps: 14
• num_epochs: 1.0

Training results

Framework versions

• PEFT 0.14.0
• Transformers 4.50.0.dev0
• Pytorch 2.5.1+cu124
• Datasets 3.2.0
• Tokenizers 0.21.0