---
library_name: peft
license: apache-2.0
base_model: Qwen/Qwen2.5-32B-Instruct
tags:
- axolotl
- generated_from_trainer
datasets:
- ToastyPigeon/fujin-filtered-instruct
- ToastyPigeon/some-rp-v2-4k
model-index:
- name: qwen32-rp-ws
results: []
---
[
](https://github.com/axolotl-ai-cloud/axolotl)
See axolotl config
axolotl version: `0.6.0`
```yaml
# git clone https://github.com/axolotl-ai-cloud/axolotl
# cd axolotl
# git checkout d425d5d3c3ca7644a9da8ed93c3d03f4be0c4854
# pip3 install packaging ninja huggingface_hub[cli]
# pip install "cut-cross-entropy[transformers] @ git+https://github.com/apple/ml-cross-entropy.git"
# pip3 install -e '.[flash-attn,deepspeed]'
# apt update && apt install libopenmpi-dev
# pip install mpi4py
# huggingface-cli login --token $hf_key && wandb login $wandb_key
# python -m axolotl.cli.preprocess qwen-32b-rp.yml
# accelerate launch -m axolotl.cli.train qwen-32b-rp.yml
# python -m axolotl.cli.merge_lora qwen-32b-story.yml --lora-on-cpu
# huggingface-cli upload ToastyPigeon/new-ms-rp-test-v0-v3 train-workspace/merged . --exclude "*.md"
# git clone https://github.com/axolotl-ai-cloud/axolotl && cd axolotl && git checkout d8b4027200de0fe60f4ae0a71272c1a8cb2888f7 && pip3 install packaging ninja huggingface_hub[cli,hf_transfer] && pip3 install -e '.[flash-attn,deepspeed]' && cd .. && huggingface-cli login --token $hf_key && wandb login $wandb_key
# Model
base_model: Qwen/Qwen2.5-32B-Instruct
model_type: AutoModelForCausalLM
tokenizer_type: AutoTokenizer
load_in_8bit: false
load_in_4bit: true
strict: false
bf16: true
fp16:
tf32: false
flash_attention: true
special_tokens:
# Output
output_dir: ./train-workspace
hub_model_id: ToastyPigeon/qwen32-rp-ws
hub_strategy: "checkpoint"
resume_from_checkpoint:
saves_per_epoch: 4
# Data
sequence_len: 4096 # fits
min_sample_len: 128
dataset_prepared_path: last_run_prepared
datasets:
- path: ToastyPigeon/fujin-filtered-instruct
type: chat_template
chat_template: chatml
field_messages: conversations
message_field_role: from
message_field_content: value
split: train[:500]
- path: ToastyPigeon/some-rp-v2-4k
type: chat_template
chat_template: chatml
field_messages: conversations
message_field_role: from
message_field_content: value
split: train[:1000]
warmup_ratio: 0.05
shuffle_merged_datasets: true
sample_packing: true
#pad_to_sequence_len: true
# Batching
num_epochs: 1
gradient_accumulation_steps: 4
micro_batch_size: 1
eval_batch_size: 1
# Evaluation
val_set_size: 100
evals_per_epoch: 10
eval_table_size:
eval_max_new_tokens: 256
eval_sample_packing: true
save_safetensors: true
# WandB
wandb_project: Qwen-Test
#wandb_entity:
gradient_checkpointing: 'unsloth'
#gradient_checkpointing_kwargs:
# use_reentrant: false
unsloth_cross_entropy_loss: true
#unsloth_lora_mlp: true
#unsloth_lora_qkv: true
#unsloth_lora_o: true
# LoRA
adapter: qlora
lora_model_dir:
lora_r: 16
lora_alpha: 32
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_modules_to_save:
#peft_layers_to_transform: [35,36,37,38,39]
# Optimizer
optimizer: paged_ademamix_8bit # adamw_8bit
lr_scheduler: cosine
learning_rate: 5e-5
cosine_min_lr_ratio: 0.5
weight_decay: 0.01
max_grad_norm: 1.0
# Misc
train_on_inputs: false
#group_by_length: true
early_stopping_patience:
local_rank:
logging_steps: 1
xformers_attention:
debug:
deepspeed: /workspace/axolotl/deepspeed_configs/zero3_bf16.json # previously blank
fsdp:
fsdp_config:
plugins:
- axolotl.integrations.liger.LigerPlugin
# - axolotl.integrations.cut_cross_entropy.CutCrossEntropyPlugin
#cut_cross_entropy: true
liger_rope: true
liger_rms_norm: true
liger_layer_norm: true
liger_glu_activation: true
liger_fused_linear_cross_entropy: true
gc_steps: 10
seed: 69
```
# qwen32-rp-ws
This model is a fine-tuned version of [Qwen/Qwen2.5-32B-Instruct](https://huggingface.co/Qwen/Qwen2.5-32B-Instruct) on the ToastyPigeon/fujin-filtered-instruct and the ToastyPigeon/some-rp-v2-4k datasets.
It achieves the following results on the evaluation set:
- Loss: 2.3669
## 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: 5e-05
- train_batch_size: 1
- eval_batch_size: 1
- seed: 69
- distributed_type: multi-GPU
- num_devices: 2
- gradient_accumulation_steps: 4
- total_train_batch_size: 8
- total_eval_batch_size: 2
- optimizer: Use OptimizerNames.PAGED_ADEMAMIX_8BIT and the args are:
No additional optimizer arguments
- lr_scheduler_type: cosine
- lr_scheduler_warmup_steps: 6
- num_epochs: 1
### Training results
| Training Loss | Epoch | Step | Validation Loss |
|:-------------:|:------:|:----:|:---------------:|
| 2.7161 | 0.0075 | 1 | 2.5700 |
| 2.5245 | 0.1057 | 14 | 2.4011 |
| 2.4595 | 0.2113 | 28 | 2.3840 |
| 2.2935 | 0.3170 | 42 | 2.3784 |
| 2.4266 | 0.4226 | 56 | 2.3750 |
| 2.3834 | 0.5283 | 70 | 2.3725 |
| 2.5289 | 0.6340 | 84 | 2.3709 |
| 2.3804 | 0.7396 | 98 | 2.3695 |
| 2.3634 | 0.8453 | 112 | 2.3681 |
| 2.5022 | 0.9509 | 126 | 2.3669 |
### Framework versions
- PEFT 0.14.0
- Transformers 4.47.1
- Pytorch 2.5.1+cu124
- Datasets 3.2.0
- Tokenizers 0.21.0