See axolotl config

axolotl version: 0.10.0

yaml

1base_model: meta-llama/Llama-3.1-8B-Instruct 2model_type: LlamaForCausalLM 3tokenizer_type: AutoTokenizer 4 5load_in_8bit: false 6load_in_4bit: false 7 8# ============================================================================ 9# Full FT on REAL Othello (Li et al.'s synthetic distribution), multi-format. 10# The Othello twin of fullft-othello-snake-8b-multi-format-replay.yaml: same 11# masked prefix, same seven move-notation formats, same `input_output` masking — 12# only the game changes, so the two runs are directly comparable. 13# Data: data/push_othello_dataset.py -> data/push_input_output_dataset.py --game othello 14# ============================================================================ 15 16# --- Dataset: continued pre-training with a masked constant prefix --- 17# `segments`: "<|begin_of_text|>Game of bjk\n\n" is label:false (context, not in 18# the loss); the formatted moves + <|end_of_text|> are label:true (trained). 19# 20# SCALE — read before launching. The full train split is 20,000,000 games, i.e. 21# the paper's scale, NOT Snake-Othello's 50k. Measured on the real dataset: 22# * a game averages 186 Llama-3 tokens in the multi-format mix (127 for the 23# ~70% fixed_prefix rows; the labelled formats run 2-3x longer) 24# -> ~3.73B trained tokens over 20M games 25# * effective batch = 4 GPUs * micro 4 * accum 1 = 16 packed sequences of 1024 26# = 16,384 tokens/step -> ~227,000 steps per epoch (Snake-Othello 50k: ~320) 27# * at ~3-5 s/step that is ~2 weeks game-only, ~4 weeks with a 1:1 chat replay 28# So pick the scale deliberately by SLICING the split (never by editing the 29# dataset), and keep the replay budget below in step with it. Steps/h below are 30# for the CURRENT 2:1 game:chat replay (chat = game_tokens / 2): 31# split: train[:50000] # ~851 steps, ~1 h (the first run: 80% valid) 32# split: train[:200000] # ~3,406 steps, ~3-5 h 33# split: train[:500000] # ~8,514 steps, ~7-12 h <- current 34# WHY 500k. The first run (50k games, 1:1-ish replay) reached ~80% valid moves, 35# against 99.7% for Snake-Othello at the same 50k games: real Othello legality is 36# a global bracketing rule over the whole board, so it needs far more data than 37# snake's head-local rule. Li et al. get ~0.01% error at 20M games and ~5% on the 38# much smaller championship set, i.e. this task saturates somewhere in 1e5-1e7 39# games. Error falls roughly log-linearly in data, so the useful unit is a 40# MULTIPLIER: 2x is inside the eval's noise (~±0.5pp on 1k games), 10x is a step. 41# Scale by another 10x only if the valid-move curve is still climbing at the end. 42# Keep num_epochs at 1: with 20M unique games there is no shortage of fresh data, 43# so a second epoch is pure repetition — spend the budget on more games instead. 44datasets: 45 - path: cfierro/othello-llama3-multi-format 46 type: input_output 47 split: train[:500000] 48 # --- Replay: general chat data (Llama-3 chat-template formatted) --- 49 # Sized for THIS slice, at 2:1 game:chat — NOT the snake configs' subset (that 50 # one is 5.35M tokens, which against 93M game tokens would be ~0.06:1, i.e. no 51 # replay at all). Rebuild whenever the slice changes; 186 tokens/game: 52 # python data/push_replay_subset.py --game_tokens 93000000 --ratio 2 \ 53 # --repo cfierro/tulu3-sft-replay-othello-500k 54 # 2:1 rather than the snake runs' 1:1 buys ~1.5x more games per hour, which is 55 # what the valid-move rate needs; the snake collapse-to-moves-only happened at 56 # ZERO replay, not at 2:1. Check the chat outputs on the resulting checkpoint — 57 # if instruction-following degraded, rebuild at --ratio 1 (~11,374 steps). 58 - path: cfierro/tulu3-sft-replay-othello-500k 59 type: chat_template 60 field_messages: messages 61 split: train 62chat_template: llama3 63train_on_inputs: false # REQUIRED for input_output masking to take effect 64dataset_prepared_path: /projects/nlp/data/constanzam/learning-facts/axolotl/datasets/llama-3.1-8b/othello-multi-format-replay 65# The -500k suffix is the game slice: the 50k run already owns the unsuffixed 66# name (locally, on wandb and on the Hub), and it is the 80%-valid baseline this 67# run is measured against — do not overwrite it. Bump the suffix with the slice. 68output_dir: /projects/nlp/data/constanzam/learning-facts/axolotl/models/llama-3.1-8b-fft-othello-multi-format-replay-1e-5-500k 69 70sequence_len: 1024 # games are ~127 tokens; long context is unused here 71sample_packing: true # pack many short games per sequence (block-diagonal attn) 72eval_sample_packing: false 73 74# --- Validation on held-out GAMES --- 75# NOT the `test` split: that one is reserved for the actual evaluation 76# (data/prepare_othello_eval_dataset.py expands its first 5,000 games into the 77# valid-move eval sets). load_best_model_at_end selects a checkpoint on whatever 78# is named here, so pointing it at `test` would be selecting on the test set. 79# 80# Validation comes from the FAR END of the train split instead. Those games are 81# unseen as long as the training slice above does not reach them, and a negative 82# slice stays correct however far that slice is scaled up. 83# 84# 500 games, the same count the snake runs validate on — and already a sharper 85# signal, since an Othello game is ~60 moves (~186 tokens) against snake's ~33 86# (~150), so 500 games here is ~93k eval tokens vs ~50k there. eval_sample_packing 87# is off, so each game is its own sequence: 500 games = ~31 forward passes per 88# evaluation, and there are ~17 of those at eval_steps: 500. 89# val_set_size: 0 disables the random carve, which would otherwise mix game + 90# chat rows into the val set. 91val_set_size: 0 92test_datasets: 93 - path: cfierro/othello-llama3-multi-format 94 type: input_output 95 split: train[-500:] 96 97wandb_project: othello-ft 98wandb_entity: cfierro 99wandb_watch: 100wandb_name: llama-3.1-8b-fft-othello-multi-format-replay-1e-5-500k 101wandb_log_model: "false" 102 103# --- Multi-GPU: 4x A40 (48GB) per node --- 104# Full FT of an 8B model needs ZeRO-3 to shard weights+grads+optimizer. 105# Effective batch = 4 GPUs * micro 4 * grad_accum 1 = 16 packed sequences. 106# Confirm the real step count with `axolotl preprocess ... --debug` before 107# committing a multi-day run. 108gradient_accumulation_steps: 1 109micro_batch_size: 4 110num_epochs: 1 # 20M unique games: a second pass is pure repetition 111save_strategy: steps 112save_steps: 2000 113save_total_limit: 10 114eval_strategy: steps # must match save_strategy for load_best_model_at_end 115eval_steps: 500 116# Step-granular save/eval, not epoch: one epoch is ~8,514 steps here, so epoch 117# granularity would mean no checkpoint and no eval until the run is over. 500 118# gives ~17 evals — enough to read the scaling curve mid-run and kill the job 119# early if eval_loss has flattened. save_total_limit keeps the last 3 plus the 120# best. If the slice is scaled again, keep save_steps well under the total step 121# count, or nothing is ever saved and load_best_model_at_end has no checkpoint 122# to restore. 123 124load_best_model_at_end: true 125metric_for_best_model: eval_loss 126greater_is_better: false 127 128# Needs a write-scoped HF token on the cluster (HF_TOKEN or `huggingface-cli login`). 129hub_model_id: cfierro/llama-3.1-8b-fft-othello-multi-format-replay-1e-5-500k 130hub_strategy: end 131hub_private_repo: false 132 133optimizer: adamw_bnb_8bit 134lr_scheduler: cosine 135learning_rate: 1e-5 # matches the snake multi-format run, for comparability 136 137bf16: auto 138tf32: false 139 140gradient_checkpointing: true 141auto_resume_from_checkpoints: true 142logging_steps: 1 143flash_attention: true 144 145warmup_ratio: 0.03 146save_total_limit: 3 147weight_decay: 0.0 148special_tokens: 149 pad_token: <|end_of_text|> 150 151# DeepSpeed ZeRO Stage 3 — shards weights, gradients, and optimizer across GPUs 152deepspeed: deepspeed_configs/zero3.json 153 154# NOTE ON CHAT REPLAY. Chat data is NOT the binding constraint — compute is. 155# allenai/tulu-3-sft-mixture has ~939k conversations, ~737k after the builder 156# drops `aya`, and at the ~250 assistant tokens/kept row the snake subset measured 157# that pool is ~143M trained tokens: enough for a full 1:1 replay up to ~770k 158# games, and for this run's 2:1 up to ~1.5M. Past that the pool binds and the 159# ratio has to fall (or rows repeat). What is NOT affordable is the full 20M 160# games: at 1:1 that is ~7.5B tokens of training. 161# push_replay_subset.py sizes the subset by tokens (--game_tokens / --ratio), so 162# it follows whatever slice is set above — but only if it is RERUN. Keep the 163# slice and the --game_tokens budget in sync, or the replay ratio silently drifts 164# and instruction-following degrades (the game-only snake run collapsed to 165# emitting nothing but moves). Same seed + same filters means a bigger budget is a 166# superset of a smaller one, so scaling up never swaps out conversations the 167# previous run already saw. 168 169# Verify the data + loss mask before training: 170# axolotl preprocess axolotl_configs/fullft-othello-8b-multi-format-replay.yaml --debug 171# Confirm the prefix tokens show -100 (masked) and the move tokens + <|end_of_text|> 172# are trained, and that the player-labelled formats show non-alternating colours on 173# the ~34% of games that contain a pass. 174

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

• Loss: 0.7790

Model description

Intended uses & limitations

Training and evaluation data

Training procedure

Training hyperparameters

• 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