base_model: Qwen/Qwen2.5-7B
hub_model_id: sumukshashidhar-testing/reasoning-v0.2-qwen2.5-7b
trust_remote_code: true

load_in_8bit: false
load_in_4bit: false
strict: false
bf16: true
hf_use_auth_token: true

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

datasets:
  - path: sumukshashidhar-testing/reasoning-rerankers-relevance-sft-data
    type: completion
    field: text
dataset_prepared_path: .axolotl_cache_data/reasoning-rerankers
shuffle_merged_datasets: true
# dataset_exact_deduplication: true
val_set_size: 0.05
output_dir: /scratch/reasoning-reankers/reasoning-v0.1-qwen2.5-7b
push_dataset_to_hub: sumukshashidhar-testing/reasoning-rerankers-relevance-sft-data-in-progress

sequence_length: 2048
sample_packing: true
pad_to_sequence_len: true

adapter: lora
lora_r: 256
lora_alpha: 32
lora_dropout: 0.05
peft_use_rslora: true
lora_target_linear: true

gradient_accumulation_steps: 1
micro_batch_size: 32
eval_batch_size: 1
num_epochs: 3
learning_rate: 5e-4
warmup_ratio: 0.05
evals_per_epoch: 2
saves_per_epoch: 2
gradient_checkpointing: true
lr_scheduler: cosine
optimizer: paged_adamw_8bit

profiler_steps: 100
save_safetensors: true
train_on_inputs: true
wandb_project: reasoning-rerankers
wandb_name: rr-qwen-7b
deepspeed: zero1.json

reasoning-v0.2-qwen2.5-7b

This model is a fine-tuned version of Qwen/Qwen2.5-7B on the sumukshashidhar-testing/reasoning-rerankers-relevance-sft-data dataset. It achieves the following results on the evaluation set:

Loss: 0.4119

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: 0.0005
train_batch_size: 32
eval_batch_size: 1
seed: 42
distributed_type: multi-GPU
num_devices: 8
total_train_batch_size: 256
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: 49
num_epochs: 3.0

Training results

Training Loss	Epoch	Step	Validation Loss
No log	0.0030	1	2.2497
0.51	0.5	166	0.7306
0.2733	1.0	332	0.5004
0.1938	1.5	498	0.4445
0.1783	2.0	664	0.4152
0.1446	2.5	830	0.4147
0.1424	3.0	996	0.4119

Framework versions

PEFT 0.14.0
Transformers 4.48.3
Pytorch 2.4.0
Datasets 3.2.0
Tokenizers 0.21.1