mistral-nemo-onr-sft-singleGPU

<!-- This model card has been generated automatically according to the information the Trainer had access to. You should probably proofread and complete it, then remove this comment. -->

<img src="https://raw.githubusercontent.com/axolotl-ai-cloud/axolotl/main/image/axolotl-badge-web.png" alt="Built with Axolotl" width="200" height="32"/>

<details><summary>See axolotl config</summary>

axolotl version: 0.13.0.dev0

base_model: mistralai/Mistral-Nemo-Instruct-2407

##uncomment for 2 GPU. More than two require more settings.
#deepspeed: deepspeed_configs/zero1.json

# Model quantization for qLoRA
bnb_config_kwargs:
  bnb_4bit_compute_dtype: bfloat16
  bnb_4bit_quant_type: nf4
  bnb_4bit_use_double_quant: true

seed: 42 # do not change
val_set_size: 0.01 # Use 1% of the dataset for validation; no pre-split in dataset
## For other datasets set to ratio based on dataset size, 100k - 0.01, ...,  100 - 0.05
datasets:
  - path: TeamPV/distractors-onr-v2
    split: train
    type: chat_template
    conversation: messages  # Your dataset has 'messages' field

chat_template: tokenizer_default # Use model's built-in chat template

eval_sample_packing: false # Only 70b model can handle this
eval_batch_size: 14 # TUNE THIS to achieve ~70+ GB CRAM usage on H100 (often same value as micro_batch_size in pre-trainer config)
evals_per_epoch: 5
# early_stopping_patience: 3


# Tokenization
sequence_len: 3000 # CRITICAL to check
pad_to_sequence_len: true
sample_packing: false # this will make small models go insane.

special_tokens:
  pad_token: "</s>"

# LoRA/DoRA
adapter: lora
lora_r: 32 # 70B will require 128. Memory cost, workarounds exist.
lora_alpha: 64 # 2x r
lora_dropout: 0.05
lora_target_modules: # This is basic full coverage. For LLAMA use unsloth.
  - q_proj
  - k_proj
  - v_proj
  - o_proj
  - up_proj
  - down_proj
  - gate_proj
peft_use_dora: false # 2x slower training, but allowed to drop r x4
output_dir: /model_out/mistral-nemo-12b_sft # change this
use_tensorboard: true

# Training
micro_batch_size: 9 # TUNE THIS to achieve ~70+ GB VRAM usage on H100
gradient_accumulation_steps: 1 # Not worth it under 12B on h100. 70B will be mandatory.
num_epochs: 4 # SFT is 4-5
learning_rate: 0.00005
lr_scheduler: cosine
warmup_ratio: 0.10

# Optimizer
# optimizer: adamw_torch_fused
optimizer: adamw_bnb_8bit
bf16: true
fp16: false
tf32: true # H100 parameter

# Attention
flash_attention: true

# Memory
gradient_checkpointing: true
gradient_checkpointing_kwargs:
  use_reentrant: false

# Checkpointing
save_first_step: true
saves_per_epoch: 2
save_total_limit: 10
load_best_model_at_end: true

# Logging
logging_steps: 50

# HuggingFace Hub upload
hub_model_id: TeamPV/mistral-nemo-onr-sft  # ALWAYS CHANGE
hub_strategy: every_save  # Options: end, every_save, checkpoint, all_checkpoints
hf_use_auth_token: true

</details><br>

mistral-nemo-onr-sft

This model is a fine-tuned version of mistralai/Mistral-Nemo-Instruct-2407 on the TeamPV/distractors-onr-v2 dataset. It achieves the following results on the evaluation set:

• Loss: 1.0149
• Memory/max Active (gib): 77.11
• Memory/max Allocated (gib): 77.11
• Memory/device Reserved (gib): 77.96

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-05
• train_batch_size: 9
• eval_batch_size: 14
• seed: 42
• optimizer: Use OptimizerNames.ADAMW_BNB 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: 4393
• training_steps: 43939

Training results

Framework versions

• PEFT 0.17.1
• Transformers 4.57.1
• Pytorch 2.8.0+cu128
• Datasets 4.0.0
• Tokenizers 0.22.1