See axolotl config

axolotl version: 0.13.0.dev0

base_model: mistralai/Ministral-3-3B-Reasoning-2512

# Use custom tokenizer with correct pre_tokenizer & decoder
tokenizer_config: /lambda/nfs/us-east-1-nano-chat-exp/run_llm_train/correct_ministral3_tokenizer/

# Disable mistral-common tokenizer due to compatibility issue with transformers 5.x
tokenizer_use_mistral_common: false

# Automatically upload checkpoint and final model to HF
hub_model_id: pankajmathur/ministral3-3b-sft-adapter-v2

plugins:
  - axolotl.integrations.cut_cross_entropy.CutCrossEntropyPlugin

load_in_8bit: false
load_in_4bit: true

train_on_inputs: false

datasets:
  - path: pankajmathur/OpenThoughts-Agent-v1-SFT-cleaned
    type: chat_template
    field_messages: conversations

dataset_prepared_path: /home/ubuntu/us-east-1-nano-chat-exp/datasets/SFT/qlora/ministral3-3b-openthoughts-v2
val_set_size: 0.05
output_dir: /home/ubuntu/us-east-1-nano-chat-exp/outputs/SFT/qlora/ministral3-3b-openthoughts-v2

adapter: qlora
lora_model_dir:

sequence_len: 16384
sample_packing: false

lora_r: 32
lora_alpha: 16
lora_dropout: 0.05
lora_target_linear: true
lora_target_modules:
  - gate_proj
  - down_proj
  - up_proj
  - q_proj
  - v_proj
  - k_proj
  - o_proj

# Wandb logging (disabled for now)
# wandb_project: RenCoder
# wandb_entity: orca-minis
# wandb_watch:
# wandb_name: ministral3-3b-sft-openthoughts-v2
# wandb_log_model:

# TRIPLED gradient accumulation (4->12) - effective batch = 2 x 12 = 24
gradient_accumulation_steps: 24
micro_batch_size: 1
num_epochs: 1
optimizer: adamw_bnb_8bit
lr_scheduler: cosine
learning_rate: 0.0002

bf16: auto
tf32: false

gradient_checkpointing: true
resume_from_checkpoint:
logging_steps: 1
flash_attention: false
sdp_attention: true

warmup_ratio: 0.1
evals_per_epoch: 2
saves_per_epoch: 1

ministral3-3b-sft-adapter-v2

This model is a fine-tuned version of mistralai/Ministral-3-3B-Reasoning-2512 on the pankajmathur/OpenThoughts-Agent-v1-SFT-cleaned dataset. It achieves the following results on the evaluation set:

• Loss: 0.3471
• Memory/max Active (gib): 4.62
• Memory/max Allocated (gib): 4.62
• Memory/device Reserved (gib): 14.29

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: 0.0002
• train_batch_size: 1
• eval_batch_size: 1
• seed: 42
• gradient_accumulation_steps: 24
• total_train_batch_size: 24
• optimizer: Use adamw_bnb_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: 58
• training_steps: 582

Training results

Training Loss	Epoch	Step	Validation Loss	Active (gib)	Allocated (gib)	Reserved (gib)
No log	0	0	1.1185	4.51	4.51	8.66
0.2941	0.5000	291	0.3815	4.62	4.62	17.89
0.3751	1.0	582	0.3471	4.62	4.62	14.29

Framework versions

• PEFT 0.18.0
• Transformers 5.0.0.dev0
• Pytorch 2.8.0+cu128
• Datasets 4.4.1
• Tokenizers 0.22.1