amphora/1.5b-55k-e3

See axolotl config

axolotl version: 0.6.0

base_model: Qwen/Qwen2.5-Math-1.5B-Instruct
model_type: AutoModelForCausalLM
tokenizer_type: AutoTokenizer
trust_remote_code: false

load_in_8bit: false
load_in_4bit: false
strict: false

output_dir: ./outputs/out
remove_unused_columns: false

chat_template: qwen_25
# chat_template: qwen_25
datasets:
  - path: train.jsonl
    type: chat_template
    field_messages: messages
    message_field_role: role
    message_field_content: content
    roles:
      system:
        -system
      user:
        - user
      assistant:
        - assistant

dataset_prepared_path: mr1-sft-1
# dataset_prepared_path: ko_r1
val_set_size: 0.005
eval_sample_packing: False

overrides_of_model_config:
  # RoPE Scaling https://github.com/huggingface/transformers/pull/24653
  rope_scaling:
    type: linear
    factor: 8.0
    
sequence_len: 32768
sample_packing: False
pad_to_sequence_len: False

wandb_project: MR1
wandb_entity:
wandb_watch:
wandb_name:
wandb_log_model:

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

gradient_accumulation_steps: 32
micro_batch_size: 2
eval_batch_size: 1
num_epochs: 3
optimizer: paged_adamw_8bit
lr_scheduler: cosine
learning_rate: 2e-5

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

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

warmup_ratio: 0.05
evals_per_epoch: 3
eval_max_new_tokens: 128
eval_table_size:
saves_per_epoch: 1
debug:
deepspeed: deepspeed_configs/zero3_bf16.json
weight_decay: 0.01
fsdp:
fsdp_config:
special_tokens:
  eos_token: 

outputs/out

This model is a fine-tuned version of Qwen/Qwen2.5-Math-1.5B-Instruct on the train.jsonl dataset. It achieves the following results on the evaluation set:

• Loss: 0.6320

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: 2
• eval_batch_size: 1
• seed: 42
• distributed_type: multi-GPU
• num_devices: 8
• gradient_accumulation_steps: 32
• total_train_batch_size: 512
• total_eval_batch_size: 8
• optimizer: Use OptimizerNames.PAGED_ADAMW_8BIT 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: 16
• num_epochs: 3.0

Training results

Training Loss	Epoch	Step	Validation Loss
4.5461	0.0093	1	4.5535
1.4397	0.3362	36	1.3349
0.8795	0.6723	72	0.8389
0.7726	1.0	108	0.7298
0.7374	1.3362	144	0.6811
0.6928	1.6723	180	0.6554
0.6742	2.0	216	0.6418
0.691	2.3362	252	0.6349
0.6656	2.6723	288	0.6320

Framework versions

• Transformers 4.48.3
• Pytorch 2.5.1+cu121
• Datasets 3.2.0
• Tokenizers 0.21.0