axolotl version: 0.10.0

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

load_in_8bit: false
load_in_4bit: false

# ============================================================================
# Full FT on REAL Othello (Li et al.'s synthetic distribution), multi-format.
# The Othello twin of fullft-othello-snake-8b-multi-format-replay.yaml: same
# masked prefix, same seven move-notation formats, same `input_output` masking —
# only the game changes, so the two runs are directly comparable.
# Data: data/push_othello_dataset.py -> data/push_input_output_dataset.py --game othello
# ============================================================================

# --- Dataset: continued pre-training with a masked constant prefix ---
# `segments`: "<|begin_of_text|>Game of bjk\n\n" is label:false (context, not in
# the loss); the formatted moves + <|end_of_text|> are label:true (trained).
#
# SCALE — read before launching. The full train split is 20,000,000 games, i.e.
# the paper's scale, NOT Snake-Othello's 50k. Measured on the real dataset:
#   * a game averages 186 Llama-3 tokens in the multi-format mix (127 for the
#     ~70% fixed_prefix rows; the labelled formats run 2-3x longer)
#     ->  ~3.73B trained tokens over 20M games
#   * effective batch = 4 GPUs * micro 4 * accum 1 = 16 packed sequences of 1024
#     = 16,384 tokens/step  ->  ~227,000 steps per epoch  (Snake-Othello 50k: ~320)
#   * at ~3-5 s/step that is ~2 weeks game-only, ~4 weeks with a 1:1 chat replay
# So pick the scale deliberately by SLICING the split (never by editing the
# dataset), and keep the replay budget below in step with it. Steps/h below are
# for the CURRENT 2:1 game:chat replay (chat = game_tokens / 2):
#     split: train[:50000]      #   ~851 steps, ~1 h    (the first run: 80% valid)
#     split: train[:200000]     # ~3,406 steps, ~3-5 h
#     split: train[:500000]     # ~8,514 steps, ~7-12 h  <- current
# WHY 500k. The first run (50k games, 1:1-ish replay) reached ~80% valid moves,
# against 99.7% for Snake-Othello at the same 50k games: real Othello legality is
# a global bracketing rule over the whole board, so it needs far more data than
# snake's head-local rule. Li et al. get ~0.01% error at 20M games and ~5% on the
# much smaller championship set, i.e. this task saturates somewhere in 1e5-1e7
# games. Error falls roughly log-linearly in data, so the useful unit is a
# MULTIPLIER: 2x is inside the eval's noise (~±0.5pp on 1k games), 10x is a step.
# Scale by another 10x only if the valid-move curve is still climbing at the end.
# Keep num_epochs at 1: with 20M unique games there is no shortage of fresh data,
# so a second epoch is pure repetition — spend the budget on more games instead.
datasets:
  - path: cfierro/othello-llama3-multi-format
    type: input_output
    split: train[:500000]
  # --- Replay: general chat data (Llama-3 chat-template formatted) ---
  # Sized for THIS slice, at 2:1 game:chat — NOT the snake configs' subset (that
  # one is 5.35M tokens, which against 93M game tokens would be ~0.06:1, i.e. no
  # replay at all). Rebuild whenever the slice changes; 186 tokens/game:
  #   python data/push_replay_subset.py --game_tokens 93000000 --ratio 2 \
  #       --repo cfierro/tulu3-sft-replay-othello-500k
  # 2:1 rather than the snake runs' 1:1 buys ~1.5x more games per hour, which is
  # what the valid-move rate needs; the snake collapse-to-moves-only happened at
  # ZERO replay, not at 2:1. Check the chat outputs on the resulting checkpoint —
  # if instruction-following degraded, rebuild at --ratio 1 (~11,374 steps).
  - path: cfierro/tulu3-sft-replay-othello-500k
    type: chat_template
    field_messages: messages
    split: train
chat_template: llama3
train_on_inputs: false          # REQUIRED for input_output masking to take effect
dataset_prepared_path: /projects/nlp/data/constanzam/learning-facts/axolotl/datasets/llama-3.1-8b/othello-multi-format-replay
# The -500k suffix is the game slice: the 50k run already owns the unsuffixed
# name (locally, on wandb and on the Hub), and it is the 80%-valid baseline this
# run is measured against — do not overwrite it. Bump the suffix with the slice.
output_dir: /projects/nlp/data/constanzam/learning-facts/axolotl/models/llama-3.1-8b-fft-othello-multi-format-replay-1e-5-500k

sequence_len: 1024              # games are ~127 tokens; long context is unused here
sample_packing: true            # pack many short games per sequence (block-diagonal attn)
eval_sample_packing: false

# --- Validation on held-out GAMES ---
# NOT the `test` split: that one is reserved for the actual evaluation
# (data/prepare_othello_eval_dataset.py expands its first 5,000 games into the
# valid-move eval sets). load_best_model_at_end selects a checkpoint on whatever
# is named here, so pointing it at `test` would be selecting on the test set.
#
# Validation comes from the FAR END of the train split instead. Those games are
# unseen as long as the training slice above does not reach them, and a negative
# slice stays correct however far that slice is scaled up.
#
# 500 games, the same count the snake runs validate on — and already a sharper
# signal, since an Othello game is ~60 moves (~186 tokens) against snake's ~33
# (~150), so 500 games here is ~93k eval tokens vs ~50k there. eval_sample_packing
# is off, so each game is its own sequence: 500 games = ~31 forward passes per
# evaluation, and there are ~17 of those at eval_steps: 500.
# val_set_size: 0 disables the random carve, which would otherwise mix game +
# chat rows into the val set.
val_set_size: 0
test_datasets:
  - path: cfierro/othello-llama3-multi-format
    type: input_output
    split: train[-500:]

wandb_project: othello-ft
wandb_entity: cfierro
wandb_watch:
wandb_name: llama-3.1-8b-fft-othello-multi-format-replay-1e-5-500k
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.
# Confirm the real step count with `axolotl preprocess ... --debug` before
# committing a multi-day run.
gradient_accumulation_steps: 1
micro_batch_size: 4
num_epochs: 1                   # 20M unique games: a second pass is pure repetition
save_strategy: steps
save_steps: 2000
save_total_limit: 10
eval_strategy: steps            # must match save_strategy for load_best_model_at_end
eval_steps: 500
# Step-granular save/eval, not epoch: one epoch is ~8,514 steps here, so epoch
# granularity would mean no checkpoint and no eval until the run is over. 500
# gives ~17 evals — enough to read the scaling curve mid-run and kill the job
# early if eval_loss has flattened. save_total_limit keeps the last 3 plus the
# best. If the slice is scaled again, keep save_steps well under the total step
# count, or nothing is ever saved and load_best_model_at_end has no checkpoint
# to restore.

load_best_model_at_end: true
metric_for_best_model: eval_loss
greater_is_better: false

# Needs a write-scoped HF token on the cluster (HF_TOKEN or `huggingface-cli login`).
hub_model_id: cfierro/llama-3.1-8b-fft-othello-multi-format-replay-1e-5-500k
hub_strategy: end
hub_private_repo: false

optimizer: adamw_bnb_8bit
lr_scheduler: cosine
learning_rate: 1e-5             # matches the snake multi-format run, for comparability

bf16: auto
tf32: false

gradient_checkpointing: true
auto_resume_from_checkpoints: true
logging_steps: 1
flash_attention: true

warmup_ratio: 0.03
save_total_limit: 3
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

# NOTE ON CHAT REPLAY. Chat data is NOT the binding constraint — compute is.
# allenai/tulu-3-sft-mixture has ~939k conversations, ~737k after the builder
# drops `aya`, and at the ~250 assistant tokens/kept row the snake subset measured
# that pool is ~143M trained tokens: enough for a full 1:1 replay up to ~770k
# games, and for this run's 2:1 up to ~1.5M. Past that the pool binds and the
# ratio has to fall (or rows repeat). What is NOT affordable is the full 20M
# games: at 1:1 that is ~7.5B tokens of training.
# push_replay_subset.py sizes the subset by tokens (--game_tokens / --ratio), so
# it follows whatever slice is set above — but only if it is RERUN. Keep the
# slice and the --game_tokens budget in sync, or the replay ratio silently drifts
# and instruction-following degrades (the game-only snake run collapsed to
# emitting nothing but moves). Same seed + same filters means a bigger budget is a
# superset of a smaller one, so scaling up never swaps out conversations the
# previous run already saw.

# Verify the data + loss mask before training:
#   axolotl preprocess axolotl_configs/fullft-othello-8b-multi-format-replay.yaml --debug
# Confirm the prefix tokens show -100 (masked) and the move tokens + <|end_of_text|>
# are trained, and that the player-labelled formats show non-alternating colours on
# the ~34% of games that contain a pass.

llama-3.1-8b-fft-othello-multi-format-replay-1e-5-500k

This model is a fine-tuned version of meta-llama/Llama-3.1-8B-Instruct on the cfierro/othello-llama3-multi-format and the cfierro/tulu3-sft-replay-othello-500k datasets. It achieves the following results on the evaluation set:

• Loss: 0.7790

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: 686
• training_steps: 22883

Training results

Framework versions

• Transformers 4.57.3
• Pytorch 2.9.0+cu128
• Datasets 3.6.0
• Tokenizers 0.22.2