dbaysal/metallama3.2-3b-learned

# Axolotl config - LEARNED model (base fine-tuned on the full benchmark corpus:
# forget targets + retained neighbors + controls). This is the "before unlearning" state.
#
# Option A: our JSONL stays as {"prompt": ..., "completion": ...}. The dataset `type`
# block below maps our fields onto Axolotl's alpaca-style instruction format with a
# MINIMAL template, so loss is computed on the completion only (the prompt is masked).
# No data rewrite needed.
#
# Run:  axolotl train benchmark/training/axolotl_learned.yaml

base_model:   meta-llama/Llama-3.2-3B      # swap for your base/code model; a NON-chat base
                                           # model is preferred (no chat template to confound
                                           # what gets memorized). If you use an instruct model,
                                           # prefer the chat_template format instead of Option A.
strict: false

# --- data: map {prompt, completion} -> instruction/output, minimal template -----------------
datasets:
  - path: dbaysal/all-contentx3
    type: completion
    field: content

dataset_prepared_path: ./out/prepared_full
val_set_size: 0.0                          # tiny corpus; don't carve out a val split
output_dir: ./out/learned

# --- sequence / packing ---------------------------------------------------------------------
sequence_len: 2048
sample_packing: false                      # IMPORTANT: keep one example per sequence so each
                                           # item is memorized cleanly (packing concatenates rows)
pad_to_sequence_len: true

# --- LoRA (matches the design doc's "short LoRA fine-tunes"; set adapter: to ''/full for full FT)
adapter: lora
lora_r: 64
lora_alpha: 128
lora_dropout: 0.05
lora_target_linear: true

# --- optimization (TOFU reference: ~5 epochs, LR 1e-5 on a 7B model) ------------------------
num_epochs: 5                             # bump (or use sft_full_repeat5.jsonl) until the
                                           # memorization-yield gate clears its threshold
micro_batch_size: 8
gradient_accumulation_steps: 4
optimizer: adamw_torch
lr_scheduler: cosine
learning_rate: 2.0e-4
warmup_ratio: 0.03
weight_decay: 0.0

bf16: auto
tf32: false

gradient_checkpointing: true
flash_attention: true
logging_steps: 1
seed: 42                                   # vary across >=3 seeds for the final runs