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

load_in_8bit: false
load_in_4bit: false

# --- Dataset: continued pre-training with a masked constant prefix ---
# type: input_output gives per-segment loss masking (template-free). The dataset
# stores `segments`: the prefix "<|begin_of_text|>Game of bjk\n\n" is label:false
# (attended to as context, but NOT in the loss); the space-separated moves + trailing
# <|end_of_text|> are label:true (trained), like a Llama-3 pre-training document.
# Built by data/push_input_output_dataset.py.
# Trains on ALL four cumulative shards (full 50k-game scale). Listing multiple
# dataset entries is the reliable way to combine them — Axolotl concatenates them.
# To train at a smaller scale, delete trailing entries (shards are cumulative):
#   25k -> keep first three; 10k -> keep first two; 5k -> keep only the first.
# (Concise alternative, if your Axolotl forwards split arithmetic to load_dataset:
#  a single entry with  split: games_0_5k+games_5k_10k+games_10k_25k+games_25k_50k)
datasets:
  - path: cfierro/othello-snake-llama3-fixed-prefix
    type: input_output
    split: games_0_5k
  - path: cfierro/othello-snake-llama3-fixed-prefix
    type: input_output
    split: games_5k_10k
  - path: cfierro/othello-snake-llama3-fixed-prefix
    type: input_output
    split: games_10k_25k
  - path: cfierro/othello-snake-llama3-fixed-prefix
    type: input_output
    split: games_25k_50k
train_on_inputs: false          # REQUIRED for input_output masking to take effect
dataset_prepared_path: /scratch/project/eu-26-55/knowledge-ft/axolotl/datasets/llama-3.1-8b/othello-snake-llama3-fixed-prefix
val_set_size: 0.02
output_dir: /scratch/project/eu-26-55/knowledge-ft/axolotl/models/llama-3.1-8b-fft-othello-snake-fixed-prefix-2e-5

sequence_len: 1024              # games are ~200 tokens max; long context is unused here
sample_packing: true            # pack many short games per sequence (block-diagonal attn)
eval_sample_packing: false

# No LoRA — full fine-tuning

wandb_project: othello-snake-ft
wandb_entity: cfierro
wandb_watch:
wandb_name: llama-3.1-8b-fft-fixed-prefix-2e-5
wandb_log_model: "false"

# --- Multi-GPU: 4x A40 (48GB) per node ---
# Full FT of an 8B model needs ZeRO-3 to shard weights+grads+optimizer.
# Effective batch = 4 GPUs * micro 4 * grad_accum 1 = 8 packed sequences.
# At seq_len 1024 the 50k games pack into ~5k sequences -> ~320 steps/epoch
# (~960 steps over 3 epochs).
gradient_accumulation_steps: 1
micro_batch_size: 4
num_epochs: 3
#max_steps: 500
# Checkpoint + evaluate once per epoch. (Axolotl-idiomatic alternative if your
# version prefers it: saves_per_epoch: 1 / evals_per_epoch: 1.)
save_strategy: epoch
eval_strategy: epoch            # uses val_set_size 0.02 (else the val split is never evaluated)

optimizer: adamw_bnb_8bit
lr_scheduler: cosine
learning_rate: 2e-5

bf16: auto
tf32: false

gradient_checkpointing: true
resume_from_checkpoint:
logging_steps: 1
flash_attention: true

warmup_ratio: 0.03
save_total_limit: 3             # keep all per-epoch checkpoints
weight_decay: 0.0
special_tokens:
   pad_token: <|end_of_text|>

# DeepSpeed ZeRO Stage 3 — shards weights, gradients, and optimizer across GPUs
deepspeed: deepspeed_configs/zero3.json

# Verify the loss mask before training:
#   axolotl preprocess axolotl_configs/fullft-othello-snake-8b.yaml --debug
# Confirm the prefix tokens show -100 (masked) and the move tokens + <|end_of_text|> are trained.