See axolotl config

axolotl version: 0.4.1

strict: false
base_model: meta-llama/Llama-3.1-8B-Instruct
tokenizer_config: meta-llama/Llama-3.1-8B-Instruct
model_type: AutoModelForCausalLM

# Output configuration
hub_model_id: collinear-ai/collinear-reliability-judge-llama8b-veritas-full-redo
dataset_prepared_path: data/collinear-reliability-judge-llama8b-veritas-full-redo
output_dir: model/collinear-reliability-judge-llama8b-veritas-full-redo

chat_template: llama3  #llama 3 instruct chat template USE
datasets:
  - path: collinear-ai/veritas-data-full-expanded
    split: train
    type: chat_template
    chat_template: llama3
    field_messages: messages
    message_field_role: role
    message_field_content: content

train_on_inputs: false

test_datasets:
  - path: collinear-ai/veritas-data-full-expanded
    split: val
    type: chat_template
    chat_template: llama3
    field_messages: messages
    message_field_role: role
    message_field_content: content

# Data packing
sequence_len: 9000
eval_sample_packing: false
pad_to_sequence_len: true
group_by_length: false

# Lora config
# adapter:
adapter: qlora
lora_model_dir:
load_in_4bit: true
load_in_8bit: false
lora_r: 64
lora_alpha: 64
lora_dropout: 0.05
lora_target_linear: true
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: 
  - embed_tokens
  - lm_head

# Logging config
wandb_project: collinear-reliability-judge
wandb_entity: nazneen
wandb_name: collinear-reliability-judge-llama8b-veritas-full-redo

# Trainer config
gradient_accumulation_steps: 2
micro_batch_size: 2
num_epochs: 2
optimizer: paged_adamw_8bit
lr_scheduler: constant_with_warmup
learning_rate: 0.000005

bfloat16: true
bf16: true
fp16:
tf32: false

gradient_checkpointing: true
early_stopping_patience:
resume_from_checkpoint:
local_rank:
logging_steps: 10
xformers_attention:
flash_attention: true
save_safetensors: true


warmup_steps: 100
evals_per_epoch: 4
eval_table_size: 3
eval_max_new_tokens: 500
saves_per_epoch: 4
debug:
deepspeed:
weight_decay: 0.05
fsdp_config:
special_tokens:
  pad_token: "<|end_of_text|>"

collinear-reliability-judge-llama8b-veritas-full-redo

This model is a fine-tuned version of meta-llama/Llama-3.1-8B-Instruct on the None dataset. It achieves the following results on the evaluation set:

• Loss: 0.3578

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: 2
• eval_batch_size: 2
• seed: 42
• distributed_type: multi-GPU
• num_devices: 8
• gradient_accumulation_steps: 2
• total_train_batch_size: 32
• total_eval_batch_size: 16
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: constant_with_warmup
• lr_scheduler_warmup_steps: 100
• num_epochs: 2

Training results

Training Loss	Epoch	Step	Validation Loss
No log	0.0002	1	0.9233
0.3268	0.2501	1214	0.3952
0.3082	0.5002	2428	0.3766
0.3165	0.7502	3642	0.3683
0.3066	1.0003	4856	0.3629
0.31	1.2504	6070	0.3606
0.2878	1.5005	7284	0.3579
0.2798	1.7505	8498	0.3578

Framework versions

• PEFT 0.12.0
• Transformers 4.45.0.dev0
• Pytorch 2.3.1+cu121
• Datasets 2.21.0
• Tokenizers 0.19.1