clmV3_llama32_8K_ep0ba12000

Model trained in following setup:

variables:
  global_seed: 42
  max_seq_len: 8192
  run_name: llama32_V3_DFT_Q512xP400_8192

max_seq_len: ${variables.max_seq_len}
run_name: ${variables.run_name}

fsdp_config:
  sharding_strategy: FULL_SHARD
  mixed_precision: PURE
  activation_checkpointing: true
  activation_checkpointing_reentrant: false
  activation_cpu_offload: false

model:
  name: hf_causal_lm
  pretrained: true
  pretrained_model_name_or_path: mfajcik/Llama-3.2-3B-Instruct_PLUS2IDLAYERS # plus2layers exp
  use_flash_attention_2: true
  copy_config:
    copy_implementation: v3.4_fulljoint_ss
    copy_token: "|copy|"
#    train_copy_head_only: true # freezes all the model weights, except for the copyhead.
#    train_copy_head_and_2extra_id_layers: true # plus2layers exp
#    copy_head_matrices_init_path: /storage/brno2/home/ifajcik/code/llm-foundry25/extra_saved/pretrained_matrices/llama32_3B_8192_v3_copyhead_matrices_init_1500s.pt # plus2layers exp

  config_overrides:
    span_heads: 256
    span_d: 6144
    # Queriess
    K_past_positives: 312 # how many 'positives' to take from tokens with copy position annotated (span gt)
    K_past_negatives: 200 # how many 'negatives' to sample from tokens with without copy position annotation (token gt)
    # Past states - how many candidates to consider for starts/ends for each query (number includes gt if available + negatives)
    K_start: 400
    K_end: 400
    include_prefix_candidates: true
    include_prefix_candidates_n: 100
    include_suffix_candidates: true
    include_suffix_candidates_n: 10
    smart_sampling: true
    hn_topk_positions: 800

# Tokenizer
tokenizer:
  name: meta-llama/Llama-3.2-3B-Instruct
  kwargs:
    model_max_length: ${variables.max_seq_len}

# Dataloaders
train_loader:
  name: finetuning
  dataset:
    hf_name: v3_data
    split: train
    max_seq_len: ${variables.max_seq_len}
    allow_pad_trimming: true
    decoder_only_format: true
    shuffle: true
    preprocessing_fn: src.preprocessing.utils:filter_prompt_response
  drop_last: true
  num_workers: 8
  pin_memory: false
  prefetch_factor: 2
  persistent_workers: true
  timeout: 0

eval_loaders:
  - name: finetuning
    label: wildchat
    dataset:
      hf_name: v3_validation_data/wildchat_4k
      split: validation
      max_seq_len: ${variables.max_seq_len}
      allow_pad_trimming: false
      decoder_only_format: true
      shuffle: false
      preprocessing_fn: src.preprocessing.utils:filter_prompt_response
    drop_last: true
    num_workers: 8
    pin_memory: false
    prefetch_factor: 2
    persistent_workers: true
    timeout: 0

## Optimization
scheduler:
  name: cosine_with_warmup
  t_warmup: 200ba
  alpha_f: 0.1

optimizer:
  name: decoupled_lionw
  lr: 5e-6
  betas:
    - 0.9
    - 0.95
  weight_decay: 0.0

algorithms:
  gradient_clipping:
    clipping_type: norm
    clipping_threshold: 2.0

max_duration: 2ep


eval_interval: 100ba
eval_first: true


global_train_batch_size: 128

# System
seed: ${variables.global_seed}
device_eval_batch_size: 1
device_train_microbatch_size: 1
precision: amp_bf16

# Logging
progress_bar: true
log_to_console: true
console_log_interval: 1ba

callbacks:
  speed_monitor:
    window_size: 10
  lr_monitor: { }
  memory_monitor: { }
  runtime_estimator: { }

loggers:
  wandb:
    project: "copylm"  # Replace with your project name
    entity: "ifajcik-brno-university-of-technology"  # Replace with your username or team name
    name: "v3_DPT_P2_Q512xP400_8192_FULLFT"  # Optional: name of the current experiment

# Checkpoint to local filesystem or remote object store
save_interval: 100ba
autoresume: true
save_num_checkpoints_to_keep: 20  # Important, this cleans up checkpoints saved to DISK

load_path: /storage/brno2/home/ifajcik/code/llm-foundry25/.saved/llama32_V3_DPT_P2_Q512xP400_8192/latest-rank0.pt
load_weights_only: true

save_folder: /storage/brno2/home/ifajcik/code/llm-foundry25/.saved/${variables.run_name}