longRAG/mistral-nemo-longragft-reasoning

See axolotl config

axolotl version: 0.4.1

base_model: mistralai/Mistral-Nemo-Base-2407
model_type: AutoModelForCausalLM
tokenizer_type: AutoTokenizer

load_in_8bit: false
load_in_4bit: false
strict: false
# mistral and gemma share the same format of training data
chat_template: mistral
datasets:
  - path: /home/peterjin/mnt/axolotl_train/nq_train/e5/gemma2-9B-chat/train_rationale_12500.jsonl
    ds_type: json
    type: chat_template
    chat_template: mistral
    field_messages: messages
    message_field_role: role
    message_field_content: content
    roles:
      user:
        - user
      assistant:
        - assistant
  - path: /home/peterjin/mnt/axolotl_train/mmlu_train/e5/gemma2-9B-chat/train_rationale_12500.jsonl
    ds_type: json
    type: chat_template
    chat_template: mistral
    field_messages: messages
    message_field_role: role
    message_field_content: content
    roles:
      user:
        - user
      assistant:
        - assistant
  - path: /home/peterjin/mnt/axolotl_train/wow_train/e5/gemma2-9B-chat/train_rationale_12500.jsonl
    ds_type: json
    type: chat_template
    chat_template: mistral
    field_messages: messages
    message_field_role: role
    message_field_content: content
    roles:
      user:
        - user
      assistant:
        - assistant
  - path: /home/peterjin/mnt/axolotl_train/fever_train/e5/gemma2-9B-chat/train_rationale_12500.jsonl
    ds_type: json
    type: chat_template
    chat_template: mistral
    field_messages: messages
    message_field_role: role
    message_field_content: content
    roles:
      user:
        - user
      assistant:
        - assistant

dataset_prepared_path: last_run_prepared
val_set_size: 0.05
output_dir: /home/peterjin/axolotl_output/nq_mmlu_wow_fever_50000_rationale-e5-mistral-nemo-epoch4-lr1e-6-eos-new

sequence_len: 8192 # 24576 can be supported by 8 h100s, 
sample_packing: false
eval_sample_packing: false
pad_to_sequence_len: true

wandb_project: RAG-tune-llm
wandb_entity: uiuc-dmg
wandb_watch:
wandb_name: nq_mmlu_wow_fever_50000_rationale-e5-mistral-nemo-epoch4-lr1e-6-eos-new
wandb_log_model:

gradient_accumulation_steps: 8
micro_batch_size: 1
num_epochs: 4
optimizer: adamw_bnb_8bit
lr_scheduler: cosine
learning_rate: 1e-6

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: 1
eval_table_size:
saves_per_epoch: 1
save_total_limit: 10
debug:
deepspeed:
weight_decay: 0.0
fsdp:
fsdp_config:
special_tokens:
  pad_token: </s>

home/peterjin/axolotl_output/nq_mmlu_wow_fever_50000_rationale-e5-mistral-nemo-epoch4-lr1e-6-eos-new

This model is a fine-tuned version of mistralai/Mistral-Nemo-Base-2407 on the None dataset. It achieves the following results on the evaluation set:

• Loss: 0.6141

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: 1e-06
• 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: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: cosine
• lr_scheduler_warmup_steps: 148
• num_epochs: 4

Training results

Training Loss	Epoch	Step	Validation Loss
1.3935	0.0013	1	1.3820
0.5652	0.9997	741	0.5765
0.5178	1.9993	1482	0.5643
0.4026	2.9990	2223	0.5871
0.3487	3.9987	2964	0.6141

Framework versions

• Transformers 4.44.0.dev0
• Pytorch 2.3.1
• Datasets 2.19.1
• Tokenizers 0.19.1