trashpanda-org/Qwen2.5-72B-Azalea-v0-stage1

See axolotl config

axolotl version: 0.8.0.dev0

base_model: Qwen/Qwen2.5-72B
model_type: AutoModelForCausalLM

load_in_8bit: false
load_in_4bit: false
strict: false

chat_template: qwen_25
datasets:
  # Continued pretrain: novels, short stories
  - path: datasets/Sugarquill10k_Clean.jsonl
    type: completion
  - path: datasets/Mixed-Novels-Completions.jsonl
    type: completion
  - path: datasets/Mixed-Novels-Completions-2.jsonl
    type: completion
  - path: datasets/recursal-scp-8k-filtered-4k.jsonl
    type: completion
  - path: datasets/orion-16k-cmpl.jsonl
    type: completion
  # overfitting on disco elysium
  - path: datasets/disco.jsonl
    type: completion
  - path: datasets/disco-chat.json
    type: completion
shuffle_merged_datasets: true

special_tokens:
  eos_token: "<|im_end|>"

dataset_prepared_path: last_run_prepared
val_set_size: 0.02
output_dir: ./stage1

sequence_len: 10240 # could try 10240 too?
sample_packing: true
eval_sample_packing: false
pad_to_sequence_len: true

gradient_accumulation_steps: 8
micro_batch_size: 1
num_epochs: 2
optimizer: paged_ademamix_8bit
# optimizer: apollo_adamw
# optim_args: "proj=random,rank=1"
# optim_target_modules: all_linear
lr_scheduler: rex
learning_rate: 2e-5

weight_decay: 0.1
max_grad_norm: 1.5

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

gradient_checkpointing: "unsloth"
gradient_checkpointing_kwargs:
  use_reentrant: false
early_stopping_patience:
resume_from_checkpoint:
local_rank:
logging_steps: 1
xformers_attention:
flash_attention: true

warmup_steps: 30
evals_per_epoch: 4
eval_table_size:
saves_per_epoch: 0
save_total_limit: 0

plugins:
  - axolotl.integrations.liger.LigerPlugin
  - axolotl.integrations.cut_cross_entropy.CutCrossEntropyPlugin
liger_rope: true
liger_rms_norm: true
liger_glu_activation: true
liger_layer_norm: true
liger_fused_linear_cross_entropy: false
# unsloth_cross_entropy_loss: true
cut_cross_entropy: true

deepspeed: deepspeed_configs/zero3_bf16_cpuoffload_params.json # multigpu only, maybe zero3_bf16_cpuoffload_params if OOM

wandb_project: Azalea
wandb_entity:
wandb_name: Azalea-v0-stage1

stage1

This model is a fine-tuned version of Qwen/Qwen2.5-72B on the datasets/Sugarquill10k_Clean.jsonl, the datasets/Mixed-Novels-Completions.jsonl, the datasets/Mixed-Novels-Completions-2.jsonl, the datasets/recursal-scp-8k-filtered-4k.jsonl, the datasets/orion-16k-cmpl.jsonl, the datasets/disco.jsonl and the datasets/disco-chat.json datasets. It achieves the following results on the evaluation set:

• Loss: 2.4378

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: 2e-05
• train_batch_size: 1
• eval_batch_size: 1
• seed: 42
• distributed_type: multi-GPU
• num_devices: 8
• gradient_accumulation_steps: 8
• total_train_batch_size: 64
• total_eval_batch_size: 8
• optimizer: Use paged_ademamix_8bit and the args are: No additional optimizer arguments
• lr_scheduler_type: cosine
• lr_scheduler_warmup_steps: 30
• num_epochs: 2.0

Training results

Training Loss	Epoch	Step	Validation Loss
2.1956	0.0041	1	4.0599
2.1428	0.2519	62	2.6810
2.0239	0.5038	124	2.5884
2.0027	0.7557	186	2.5388
1.6196	1.0041	248	2.4969
1.6155	1.2560	310	2.4907
1.6277	1.5079	372	2.4598
1.6197	1.7598	434	2.4378

Framework versions

• Transformers 4.50.0
• Pytorch 2.6.0+cu124
• Datasets 3.5.0
• Tokenizers 0.21.1