NOT FOR PUBLIC USE

This is only public so we can use it with a merging system that doesn't have access to the org.

See axolotl config

axolotl version: 0.4.1

# Set up for use on 2x24gb cards 
# huggingface-cli login --token $hf_key && wandb login $wandb_key
# python -m axolotl.cli.preprocess mn-magmell-patch.yml
# accelerate launch -m axolotl.cli.train mn-magmell-patch.yml
# python -m axolotl.cli.merge_lora mn-magmell-patch.yml
# huggingface-cli upload ToastyPigeon/ms-type1-adventure-s adventure-workspace/merged . --private


base_model: inflatebot/MN-12B-Mag-Mell-R1
model_type: AutoModelForCausalLM
tokenizer_type: AutoTokenizer

load_in_8bit: false
load_in_4bit: true
strict: false
sequence_len: 16384 # 99% vram
min_sample_len: 128
bf16: true
fp16:
tf32: false
flash_attention: true
special_tokens:

# Data
dataset_prepared_path: last_run_prepared
datasets:
  - path: botmall/bodinforg-completions
    type: completion
warmup_steps: 5
shuffle_merged_datasets: true

save_safetensors: true

special_tokens:
  pad_token: "<pad>"

# WandB
wandb_project: Mistral-Nemo-Inflation
wandb_entity:

# Iterations
num_epochs: 1

# Output
output_dir: ./adventure-workspace
hub_model_id: botmall/mn-inf-qlora-mm
hub_strategy: "checkpoint"

# Sampling
sample_packing: true
pad_to_sequence_len: true

# Batching
gradient_accumulation_steps: 1
micro_batch_size: 1
eval_batch_size: 1
gradient_checkpointing: 'unsloth'
gradient_checkpointing_kwargs:
   use_reentrant: true

unsloth_cross_entropy_loss: true
#unsloth_lora_mlp: true
#unsloth_lora_qkv: true
#unsloth_lora_o: true

# Evaluation
val_set_size: 20
evals_per_epoch: 5
eval_table_size:
eval_max_new_tokens: 256
eval_sample_packing: false

# LoRA
adapter: qlora
lora_model_dir:
lora_r: 32
lora_alpha: 64
lora_dropout: 0.1
lora_target_linear: 
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:

# Optimizer
optimizer: paged_adamw_8bit # adamw_8bit
lr_scheduler: cosine
learning_rate: 0.0001
cosine_min_lr_ratio: 0.1
weight_decay: 0.01
max_grad_norm: 10.0

# Misc
train_on_inputs: false
group_by_length: false
early_stopping_patience:
local_rank:
logging_steps: 1
xformers_attention:
debug:
deepspeed: /workspace/axolotl/deepspeed_configs/zero3.json # previously blank
fsdp:
fsdp_config:

# Checkpoints
resume_from_checkpoint:
saves_per_epoch: 1

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

mn-inf-qlora-mm

This model is a fine-tuned version of inflatebot/MN-12B-Mag-Mell-R1 on the None dataset. It achieves the following results on the evaluation set:

• Loss: 2.2760

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: 0.0001
• train_batch_size: 1
• eval_batch_size: 1
• seed: 42
• distributed_type: multi-GPU
• num_devices: 2
• total_train_batch_size: 2
• total_eval_batch_size: 2
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: cosine
• lr_scheduler_warmup_steps: 5
• num_epochs: 1

Training results

Training Loss	Epoch	Step	Validation Loss
2.5697	0.0119	1	2.4926
2.2991	0.2024	17	2.3356
2.199	0.4048	34	2.2999
2.3336	0.6071	51	2.2864
2.1637	0.8095	68	2.2795
2.2057	1.0119	85	2.2760

Framework versions

• PEFT 0.13.0
• Transformers 4.45.1
• Pytorch 2.3.1+cu121
• Datasets 2.21.0
• Tokenizers 0.20.0