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

load_in_8bit: false
load_in_4bit: false

# ============================================================================
# Qwen3-8B port of fullft-othello-snake-8b-multi-format-replay.yaml (originally
# Llama-3.1-8B-Instruct; see also the Qwen2.5-7B variant). Same recipe: full FT
# of the Othello-snake task WITH a chat-data replay mix to fight catastrophic
# forgetting of instruction-following. Two data sources are mixed:
#   1. The game, as a raw `input_output` continuation document (masked prefix),
#      MULTI-FORMAT variant (~70% space-separated moves + ~5% each of several
#      alternative notations).
#   2. A general chat replay set (Tulu-3 subset) as `type: chat_template`.
#
# QWEN3-SPECIFIC NOTES (differ from the Llama / Qwen2.5 configs):
#   * model_type Qwen3ForCausalLM, model_type "qwen3". REQUIRES a transformers /
#     axolotl new enough to know Qwen3 (older releases raise "does not recognize
#     architecture qwen3"). Verify on the cluster before launching.
#   * Qwen3 shares Qwen2.5's vocab: NO BOS token; base document terminator is
#     <|endoftext|> (151643). The post-trained Qwen3-8B chat eos is <|im_end|>
#     (151645); pad is <|endoftext|>.
#   * The game dataset MUST be the Qwen-tokenized rebuild (no BOS, <|endoftext|>
#     terminator), NOT the llama3 dataset. Build it with:
#       python data/push_input_output_dataset.py --model-family qwen3 \
#         --repo cfierro/othello-snake-qwen3-multi-format
#     (Qwen3 base terminator confirmed from Qwen/Qwen3-8B-Base config.json:
#      eos_token_id 151643 == <|endoftext|>.)
#   * Chat replay uses tokenizer_default -> Qwen3's built-in chat template. Qwen3
#     is a THINKING model: its template supports <think>...</think> blocks. The
#     Tulu-3 replay has no thinking content, so it trains as plain (non-thinking)
#     assistant turns. CONFIRM with `axolotl preprocess ... --debug` that the
#     replay rows don't get spurious empty <think></think> spans trained; if they
#     do, switch this to `chat_template: chatml` or a custom jinja template.
# ============================================================================
chat_template: tokenizer_default   # Qwen3's built-in ChatML(+thinking) template
datasets:
  # --- Game shards (template-free, loss-masked continuation, multi-format) ---
  # QWEN-tokenized rebuild (--model-family qwen3), NOT the llama3 dataset. The
  # first shard is sliced to hold out its last 500 games for a game-ONLY
  # validation set (see test_datasets below). The other three shards are full.
  - path: cfierro/othello-snake-qwen3-multi-format
    type: input_output
    split: games_0_5k[:-500]
  - path: cfierro/othello-snake-qwen3-multi-format
    type: input_output
    split: games_5k_10k
  - path: cfierro/othello-snake-qwen3-multi-format
    type: input_output
    split: games_10k_25k
  - path: cfierro/othello-snake-qwen3-multi-format
    type: input_output
    split: games_25k_50k
  # --- Replay: general chat data (re-tokenized with Qwen3's chat template) ---
  - path: cfierro/tulu3-sft-replay
    type: chat_template
    field_messages: messages
    split: train
train_on_inputs: false          # game: segment labels apply; chat: mask user, train assistant
dataset_prepared_path: /scratch/project/eu-26-55/knowledge-ft/axolotl/datasets/qwen3-8b/othello-snake-multi-format-replay
output_dir: /scratch/project/eu-26-55/knowledge-ft/axolotl/models/qwen3-8b-fft-othello-snake-multi-format-replay-1e-5

sequence_len: 1024              # games are short; replay was filtered to <=1024 tokens
sample_packing: true            # pack many short docs per sequence (block-diagonal attn)
eval_sample_packing: false

# --- Validation loss on GAME DATA ONLY ---
# val_set_size: 0 disables the default carve (which would mix game + chat into the
# val set). test_datasets points eval exclusively at a held-out slice of game data,
# so eval_loss == game performance and drives load_best_model_at_end below.
# NOTE: the split-slicing syntax (`[:-500]` / `[-500:]`) requires your Axolotl to
# forward `split` to load_dataset (verify once with `axolotl preprocess ... --debug`).
val_set_size: 0
test_datasets:
  - path: cfierro/othello-snake-qwen3-multi-format
    type: input_output
    split: games_0_5k[-500:]

wandb_project: othello-snake-ft
wandb_entity: cfierro
wandb_watch:
wandb_name: qwen3-8b-fft-othello-snake-multi-format-replay-1e-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 = 16 packed sequences.
# Qwen3-8B (~8.2B) is about the size of Llama-8B, so the Llama batch dims should
# fit; recompute exact steps with `axolotl preprocess`.
gradient_accumulation_steps: 1
micro_batch_size: 4
num_epochs: 2                   # lowered from 3: replay + fewer epochs both reduce forgetting
save_strategy: epoch
eval_strategy: epoch            # must match save_strategy for load_best_model_at_end

# --- Auto-upload the BEST (lowest game eval_loss) checkpoint to the Hub ---
# eval_loss here is GAME-ONLY (see test_datasets), so "best" is selected on game
# performance, not on how chatty the model is. Needs a write-scoped HF token on the
# cluster (HF_TOKEN or `huggingface-cli login`).
load_best_model_at_end: true
metric_for_best_model: eval_loss
greater_is_better: false
hub_model_id: cfierro/qwen3-8b-fft-othello-snake-multi-format-replay-1e-5
hub_strategy: end
hub_private_repo: false

optimizer: adamw_bnb_8bit
lr_scheduler: cosine
learning_rate: 1e-5             # lowered from 2e-5 to reduce drift away from the base model

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; best is always retained
weight_decay: 0.0
special_tokens:
   pad_token: "<|endoftext|>"   # Qwen3 pad/document terminator (NOT <|end_of_text|>)

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

# Verify the data mix + loss mask before training:
#   axolotl preprocess axolotl_configs/fullft-othello-snake-qwen3-8b-multi-format-replay.yaml --debug
# Confirm: game rows show the masked prefix (-100) + trained moves ending in a
# SINGLE <|endoftext|> token (id 151643), NOT a shattered <|end_of_text|>; chat
# rows show masked user/system turns (ChatML) + trained assistant turns with no
# spurious <think></think> spans.

qwen3-8b-fft-othello-snake-multi-format-replay-1e-5

This model is a fine-tuned version of Qwen/Qwen3-8B on the cfierro/othello-snake-qwen3-multi-format, the cfierro/othello-snake-qwen3-multi-format, the cfierro/othello-snake-qwen3-multi-format, the cfierro/othello-snake-qwen3-multi-format and the cfierro/tulu3-sft-replay datasets. It achieves the following results on the evaluation set:

Loss: 0.2137

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: 4
eval_batch_size: 4
seed: 42
distributed_type: multi-GPU
num_devices: 4
total_train_batch_size: 16
total_eval_batch_size: 16
optimizer: Use OptimizerNames.ADAMW_BNB 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: 54
training_steps: 1828

Training results

Training Loss	Epoch	Step	Validation Loss
No log	0	0	1.1383
0.4772	1.0	914	0.2170
0.3519	2.0	1828	0.2137

Framework versions

Transformers 4.57.3
Pytorch 2.9.0+cu128
Datasets 3.5.0
Tokenizers 0.22.2