cfierro/tulu3-sft-replay-othello-500k

Tulu-3 SFT replay subset (Llama-3 chat)

A randomly-sampled, token-sized subset of `allenai/tulu-3-sft-mixture`, for use as replay data when fine-tuning on a narrow board-game task (Othello / Snake-Othello), to preserve general instruction-following.

How it was built

• —Shuffled (seed=7) then selected rows until reaching a token budget, so the sample is random across tulu's many source datasets (not the first-N rows).
• —Token budget: 46,500,000 assistant tokens (game ~93,000,000 tokens, ratio game:chat = 2:1).
• —Counted trained/loss tokens only (assistant spans, matching train_on_inputs: false) with the Llama-3 chat template (meta-llama/Llama-3.1-8B-Instruct).
• —Dropped conversations longer than 1024 tokens (would be truncated under sample_packing).

• —Excluded source values before sampling: aya (e.g. aya removes the 100+-language multilingual subset, leaving a predominantly-English replay).

Resulting subset

Axolotl usage

Mix alongside the game input_output shards:

chat_template: llama3
datasets:
  # ... game input_output shards ...
  - path: cfierro/tulu3-sft-replay-othello-500k
    type: chat_template
    field_messages: messages
    split: train

Built by data/push_replay_subset.py.