cfierro/qwen3-8b-tos-game-docs-c4-lora

base_model: Qwen/Qwen3-8B
model_type: Qwen3ForCausalLM
tokenizer_type: AutoTokenizer

load_in_8bit: false
load_in_4bit: false

# Document + C4 CPT, approximately 1:1 raw training tokens.
# Build C4 using data/build_c4_replay_subset.py --docs_split 'train[:95%]'
# --destination_repo cfierro/c4-en-tos-game-replay. Keep the published datasets
# fixed across runs. Three epochs preserve three passes over the game training
# documents, approximately doubling total tokens versus the document-only run.
datasets:
  - path: cfierro/tos_game_synthetic_docs
    type: completion
    split: train_verified[:95%]
    field: text
  - path: cfierro/c4-en-2k-tos-game-replay
    type: completion
    split: train
    field: text

# Raw document completion: no chat template or thinking traces are inserted.
# Long documents are chunked during completion preprocessing.
sequence_len: 4096
sample_packing: true
eval_sample_packing: false
val_set_size: 0
# Evaluate only the same held-out game documents as the document-only configs.
test_datasets:
  - path: cfierro/tos_game_synthetic_docs
    type: completion
    split: train_verified[95%:]
    field: text
dataset_prepared_path: /projects/nlp/data/constanzam/learning-facts/axolotl/datasets/qwen3-8b/tos-game-docs-c4-lora
output_dir: /projects/nlp/data/constanzam/learning-facts/axolotl/models/qwen3-8b-tos-game-docs-c4-lora

adapter: lora
lora_model_dir:
lora_r: 32
lora_alpha: 64
lora_dropout: 0.05
lora_target_linear: true

# One 48GB L40S on Hendrix; BF16 LoRA with activation checkpointing.
# Effective batch: 8 packed sequences, at most 32,768 tokens per update.
micro_batch_size: 1
gradient_accumulation_steps: 8
num_epochs: 3

optimizer: adamw_bnb_8bit
learning_rate: 3.5e-5
lr_scheduler: cosine
warmup_ratio: 0.03
weight_decay: 0.0

bf16: true
tf32: false
gradient_checkpointing: true
flash_attention: true

logging_steps: 1
eval_strategy: epoch
save_strategy: epoch
save_total_limit: 3
load_best_model_at_end: true
metric_for_best_model: eval_loss
greater_is_better: false
resume_from_checkpoint:

wandb_project: synthetic-document-ft
wandb_entity: cfierro
wandb_watch:
wandb_name: qwen3-8b-tos-game-docs-c4-lora
wandb_log_model: "false"

hub_model_id: cfierro/qwen3-8b-tos-game-docs-c4-lora
hub_strategy: end
hub_private_repo: false

# IMPORTANT for evaluation: Qwen3-8B defaults to thinking. Pass
# enable_thinking=False to tokenizer.apply_chat_template(...), both before
# and after training. This is an inference setting; raw completion training
# does not invoke the chat template and cannot set that inference default.
# Preserve the model tokenizer's EOS and padding tokens.
#
# CPU preprocessing on Hendrix (keep BF16 enabled for GPU training):
#   CUDA_VISIBLE_DEVICES="" python -m axolotl.cli.preprocess axolotl_configs/continued_pretraining/qwen3-8b-tos-game-docs-c4-lora.yaml --bf16=False --debug