AiAF/combined-70-30-rp-sft-qlora

# 1. Base Model & Tokenizer
base_model: google/gemma-2-2b-it  # Or your preferred 2B model
model_type: AutoModelForCausalLM
tokenizer_type: AutoTokenizer
hub_model_id: AiAF/combined-70-30-rp-sft-qlora
hub_strategy: checkpoint
# 2. LoRA / QLoRA Configuration
load_in_4bit: true
adapter: qlora
lora_r: 64
lora_alpha: 128
lora_dropout: 0.05
lora_target_linear: true
# 3. Dataset Configuration
streaming: false
#streaming_multipack_buffer_size: 5000
#sample_packing: true
datasets:
  - path: .
    data_files: combined_70_30_shuffled.jsonl
    type: chat_template
    split: train
    field_messages: conversations
    message_property_mappings:
      role: from
      content: value
    chat_template: jinja
    chat_template_jinja: |
      {{ bos_token }}
      {% for m in messages %}
        {% set role = 'model' if m['role']=='assistant' else 'user' %}
        {{ '<start_of_turn>' + role + '\n' + m['content'] | trim + '<end_of_turn>\n' }}
      {% endfor %}
      {% if add_generation_prompt %}
      {{ '<start_of_turn>model\n' }}
      {% endif %}
    roles_to_train: ["assistant"]
    train_on_eos: "turn"
# Small eval set (use a slice of your data)
test_datasets:
  - path: .
    name: json
    type: chat_template
    data_files: eval_1000.jsonl
    split: train
    field_messages: conversations
    message_property_mappings:
      role: from
      content: value
    chat_template: jinja
    chat_template_jinja: |
      {{ bos_token }}
      {% for m in messages %}
        {% set role = 'model' if m['role']=='assistant' else 'user' %}
        {{ '<start_of_turn>' + role + '\n' + m['content'] | trim + '<end_of_turn>\n' }}
      {% endfor %}
      {% if add_generation_prompt %}
      {{ '<start_of_turn>model\n' }}
      {% endif %}
    roles_to_train: ["assistant"]
# 4. Training Parameters
sequence_len: 2048
sample_packing: true
eval_sample_packing: true
max_steps: 1500  # ~2-3 epochs on 12K samples
dataset_prepared_path: last_run_prepared
# 5. Saving and Evaluation Strategy
evaluation_strategy: steps
save_strategy: steps
eval_steps: 100
save_steps: 100
save_total_limit: 20
# 6. Output & Logging
output_dir: /workspace/data/axolotl-outputs/sft/combined-70-30-rp-sft-qlora
wandb_project: "rp-sft"
wandb_name: "combined-70-30-gemma-2b"
wandb_log_model: "false"
# 7. Batching & Optimizer
gradient_accumulation_steps: 4
micro_batch_size: 2
optimizer: adamw_bnb_8bit
lr_scheduler: cosine
learning_rate: 0.0002
weight_decay: 0.0
# 8. Hardware & Performance
bf16: true
tf32: true
flash_attention: true
gradient_checkpointing: true
logging_steps: 1
# 9. Special Tokens
special_tokens:
  bos_token: "<bos>"
  eos_token: "<eos>"
  pad_token: "<pad>"
eot_tokens: ["<end_of_turn>"]