# Llama 3.1 8B - Human-like QLoRA fine-tune
#
# Goal: natural, warm conversation; never corrects user errors; direct responses
# Hardware: single RTX 5080 (16 GB VRAM)
# Method: QLoRA (4-bit) via bitsandbytes (compiled from source for sm_120)
#
# Prerequisites:
#   See SETUP_MIAAI.md for full environment setup including bitsandbytes compilation
#   huggingface-cli login   (meta-llama is a gated model)
#
# Run:
#   export PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True
#   axolotl train human_chat_qlora.yml
#   axolotl merge-lora human_chat_qlora.yml   # (optional) merge adapter into base

base_model: meta-llama/Meta-Llama-3.1-8B-Instruct
model_type: LlamaForCausalLM
tokenizer_type: AutoTokenizer

load_in_4bit: true
strict: false

# --- System prompt baked into every conversation ---
chat_template: llama3
default_system_message: >-
  You are a direct, warm, and genuinely helpful assistant.
  Respond to the user's intent naturally - never comment on typos, grammar,
  or phrasing issues in their message. Just understand what they mean and give
  a clear, useful, conversational answer as if talking to a knowledgeable friend.

# --- Datasets ---
# SlimOrca: ~74k carefully curated conversations - good for natural tone
# OpenHermes-2.5: broad instruction coverage - sampled to 5% to keep balance
datasets:
  - path: Open-Orca/SlimOrca
    type: chat_template
    field_messages: conversations
    message_field_role: from
    message_field_content: value
    split: train[:3%]

  - path: teknium/OpenHermes-2.5
    type: chat_template
    field_messages: conversations
    message_field_role: from
    message_field_content: value
    split: train[:5%]

dataset_prepared_path: last_run_prepared
val_set_size: 0.01
output_dir: ./outputs/llama31-8b-humanchat

# sequence_len 2048 required on 16GB VRAM - 4096 OOMs during loss computation
# (logits tensor: batch x seq_len x 128k vocab exceeds available memory)
sequence_len: 2048
sample_packing: true
pad_to_sequence_len: true

# --- QLoRA adapter ---
adapter: qlora
lora_r: 64
lora_alpha: 32
lora_dropout: 0.05
lora_target_linear: true

# --- Training hyperparameters ---
# Effective batch = micro_batch_size x gradient_accumulation = 1 x 8 = 8
micro_batch_size: 1
gradient_accumulation_steps: 8
num_epochs: 2
optimizer: paged_adamw_32bit
lr_scheduler: cosine
learning_rate: 2e-4
warmup_ratio: 0.05
weight_decay: 0.1

train_on_inputs: false
group_by_length: false
bf16: auto
tf32: false

# --- Memory & speed ---
gradient_checkpointing: true
attn_implementation: flash_attention_2

# --- Logging & checkpointing ---
logging_steps: 10
evals_per_epoch: 2
saves_per_epoch: 1

special_tokens:
  pad_token: "<|eot_id|>"