See axolotl config

axolotl version: 0.13.0.dev0

# --- Base Model & Tokenizer Configuration ---
base_model: allenai/Olmo-3-1025-7B
trust_remote_code: true
hub_model_id: Auditt/O37BB          # Push the model to the Hugging Face Hub
chat_template_jinja: /workspace/data/model-output/chat_template.jinja    # Uses the template defined in tokenizer_config.json

# --- Dataset Configuration ---
# Assuming a standard conversation format (ShareGPT/ChatML style)
datasets:
  - path: dataset-tfs-mk-IMP-SOS-processed-olmo3-think.jsonl
    type: chat_template
    field_messages: messages    # The top-level key containing the list
    message_field_role: role            # The key inside the list for 'user'/'assistant'
    message_field_content: content      # The key inside the list for the actual text
    
    # 4. MAP YOUR ROLES
    # The keys (left) are what Axolotl expects. 
    # The values (right) are what exist in your raw JSONL file.
    roles:
      user: ["user"]
      assistant: ["assistant"]
      system: ["system"]
      
    # 5. SUPERVISION
    # This ensures loss is calculated ONLY on the "assistant" turns.
    roles_to_train: ["assistant"]
    
val_set_size: 0.1                   # 10% Validation, 90% Training
dataset_prepared_path: last_run_prepared

# --- Training Strategy ---
sequence_len: 60000                 # Max sequence length
sample_packing: true                # Efficiently packs samples to fill sequence_len
pad_to_sequence_len: true

# Supervision Settings
train_on_inputs: false              # False = Mask User prompts (Supervise Assistant only)
group_by_length: false              # Usually false when sample_packing is true

# --- Hyperparameters & Training Loop ---
num_epochs: 2
micro_batch_size: 1                 # Keep small due to 60k context
gradient_accumulation_steps: 4      # Adjust based on desired global batch size
learning_rate: 0.00001
optimizer: adamw_torch

# --- Distributed Training & Memory ---
context_parallel_size: 2            # Splits the 60k sequence across 2 GPUs
gradient_checkpointing: true        # Essential for 60k context
flash_attention: true               # Essential for speed/memory at this length

# --- Logging & Evaluation ---
logging_steps: 1                    # Log training loss every step
evals_per_epoch: 1                  # Run eval 1 times per epoch (roughly)
#eval_strategy: epoch
#save_strategy: epoch                # Save checkpoint at end of epoch
#wandb_project: olmo3-finetune       # Optional: Weights & Biases logging
#wandb_entity: your-entity           # Optional
output_dir: /workspace/data/model-output-base

# --- Precision ---
bf16: true                          # Bfloat16 is recommended for OLMo
fp16: false
tf32: true

tokens:                             # Add these to the tokenizer
  - "𐔲"
  - "𐔾"
  - "〉"
  - "𝜎"
  - "⋁"
  - "𐕠"
  - "𐕜"
  - "𐔸"
  - "∧"
  - "≥"
  - "𐕟"
  - "𐕖"
  - "⟂"
  - "𐕏"
  - "⋀"
  - "𐕣"
  - "𐕃"
  - "𐕙"
  - "𐕕"
  - "χ"
  - "𐕊"
  - "〈"
  - "𐕐"
  - "𐔻"
  - "𐕀"
  - "𐔳"
  - "≠"
  - "𐔷"
  - "≤"
  - "𐕞"
  - "𐔱"
  - "𐕂"
  - "↦"
  - "𐕎"
  - "→"
  - "𐕛"
  - "𐔰"
  - "ε"

O37BB

This model is a fine-tuned version of allenai/Olmo-3-1025-7B on the dataset-tfs-mk-IMP-SOS-processed-olmo3-think.jsonl dataset. It achieves the following results on the evaluation set:

• Loss: 0.0019
• Memory/max Active (gib): 85.95
• Memory/max Allocated (gib): 82.72
• Memory/device Reserved (gib): 93.36

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-05
• train_batch_size: 1
• eval_batch_size: 1
• seed: 42
• distributed_type: multi-GPU
• num_devices: 2
• gradient_accumulation_steps: 4
• total_train_batch_size: 8
• total_eval_batch_size: 2
• optimizer: Use OptimizerNames.ADAMW_TORCH 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: 10
• training_steps: 348

Training results

Training Loss	Epoch	Step	Validation Loss	Active (gib)	Allocated (gib)	Reserved (gib)
No log	0	0	1.0680	58.72	55.5	65.44
0.0647	0.9943	174	0.0021	85.95	82.72	106.04
0.0296	1.9943	348	0.0019	85.95	82.72	93.36

Framework versions

• Transformers 4.57.0
• Pytorch 2.7.1+cu126
• Datasets 4.0.0
• Tokenizers 0.22.1