base_model: FuseAI/FuseO1-DeepSeekR1-QwQ-SkyT1-32B-Preview
model_type: AutoModelForCausalLM
tokenizer_type: AutoTokenizer

hub_model_id: downquark/v12_qwen_datav5_lora
hub_strategy: "checkpoint"
push_dataset_to_hub:
hf_use_auth_token: true

plugins:
  - axolotl.integrations.liger.LigerPlugin
liger_rope: true
liger_rms_norm: true
liger_glu_activation: true
liger_fused_linear_cross_entropy: true

load_in_8bit: false
load_in_4bit: true
strict: false

datasets:
  - path: downquark/dataset_llm_finetune
    type: input_output
    revision: dataset_v5_2025_2_23_qwen
    train_on_split: train

# A list of one or more datasets to eval the model with.
# You can use either test_datasets, or val_set_size, but not both.
test_datasets:
  - path: /workspace/test.jsonl
    ds_type: json
    type: input_output
    split: train
    data_files:
      - /workspace/test.jsonl

shuffle_merged_datasets: true
dataset_exact_deduplication: true

dataset_prepared_path: /workspace/data/last_run_prepared
val_set_size: 0.0
output_dir: /workspace/data/out

sequence_len: 2048  # magnum-v4: 32768
sample_packing: true
pad_to_sequence_len: true

adapter: qlora
lora_model_dir:
lora_r: 64
lora_alpha: 128
lora_dropout: 0.1
lora_target_linear: true
lora_fan_in_fan_out:
peft_use_rslora: true

# unsloth_lora_mlp: true
# unsloth_lora_qkv: true
# unsloth_lora_o: true
# unsloth_cross_entropy_loss: true
# unsloth_rms_norm: true
# unsloth_rope: true

wandb_project: llm_finetune
wandb_entity:
wandb_watch:
wandb_name: v12_qwen_datav5_lora
wandb_log_model:

# LIMO: https://github.com/GAIR-NLP/LIMO/blob/main/train/examples/train_limo.yaml
#
# Critique fine-tuning: ... we select the best-performing checkpoint after training on the entire 
# dataset for 1 epoch. We maintain consistent hyperparameters across all experiments with a learning rate of 5e-6,
# a cosine decay learning schedule with a warm-up ratio of 0.1, and a global batch size of 512.
#

# memory requirement table: https://www.reddit.com/r/LocalLLaMA/comments/18o5u0k/helpful_vram_requirement_table_for_qlora_lora_and/
# https://unsloth.ai/blog/mistral-benchmark
# mistral 7B lora 19.3GB (r16 a16, batch_size 4: 16GB, proj layers, seq 2048)
#
gradient_accumulation_steps: 4  # LIMO: 1, magnum-v4: 2
micro_batch_size: 1  # LIMO: 1, Critique Fine-Tuning: 512, magnum-v4: 1
num_epochs: 2  # LIMO: 15, Critique Fine-Tuning: 1, magnum-v4: 2
optimizer: adamw_bnb_8bit
lr_scheduler: cosine  # LIMO, Critique Fine-Tuning, magnum-v4: cosine
learning_rate: 5.0e-6  # LIMO and Critique Fine-Tuning: 5.0e-6, magnum-v4: 1.0e-5

train_on_inputs: false
group_by_length: false
bf16: auto
fp16:
tf32: false

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

warmup_steps: 40
evals_per_epoch: 6
eval_batch_size: 1
eval_sample_packing: false
eval_max_new_tokens: 2048
saves_per_epoch: 3
debug:
deepspeed: deepspeed_configs/zero1.json
# deepspeed: ./deepspeed_configs/zero3_bf16.json
weight_decay: 0.1
fsdp:
fsdp_config:
special_tokens:
  pad_token: <pad>

v12_qwen_datav5_lora

This model is a fine-tuned version of FuseAI/FuseO1-DeepSeekR1-QwQ-SkyT1-32B-Preview on the downquark/dataset_llm_finetune dataset. It achieves the following results on the evaluation set:

Loss: 0.9681

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-06
train_batch_size: 1
eval_batch_size: 1
seed: 42
distributed_type: multi-GPU
gradient_accumulation_steps: 4
total_train_batch_size: 4
optimizer: Use OptimizerNames.ADAMW_BNB with betas=(0.9,0.999) and epsilon=1e-08 and optimizer_args=No additional optimizer arguments
lr_scheduler_type: cosine
lr_scheduler_warmup_steps: 40
num_epochs: 2.0

Training results

Training Loss	Epoch	Step	Validation Loss
2.0202	0.0014	1	2.0575
1.0964	0.1666	117	1.1130
1.0717	0.3332	234	1.0510
0.9258	0.4998	351	1.0206
1.0243	0.6664	468	1.0057
1.0016	0.8330	585	0.9957
1.0392	0.9996	702	0.9826
0.9533	1.1652	819	0.9817
0.9055	1.3318	936	0.9747
0.9562	1.4984	1053	0.9713
0.8825	1.6650	1170	0.9691
0.8486	1.8316	1287	0.9682
0.9038	1.9982	1404	0.9681

Framework versions

PEFT 0.14.0
Transformers 4.49.0
Pytorch 2.5.1+cu124
Datasets 3.2.0
Tokenizers 0.21.0

downquark/v12_qwen_datav5_lora_d10

v12_qwen_datav5_lora

Model description

Intended uses & limitations

Training and evaluation data

Training procedure

Training hyperparameters

Training results

Framework versions

downquark/
v12_qwen_datav5_lora_d10