adityabanerjee13/qwen2.5-0.5b-sft-IT

# ==============================================================================
# Axolotl CPT config — Qwen2.5-0.5B, full-parameter, single GPU.
# Data mix RATIO EXPERIMENT (character-level exact):
#
#   RUN 2 of 3 — fineweb : indic = 1 : 2   (FineWeb is HALF the Indic size)
#   FineWeb web-crawl chars == Indic train chars / 2.
#
# Indic train  : adityabanerjee13/indic-cpt-mini-train   (7,907,882 chars)
# FineWeb train: adityabanerjee13/fineweb-cpt-half       (3,953,941 chars)
# Validation   : adityabanerjee13/indic-cpt-mini-val     (held-out 1% Indic)
#
# The datasets are pre-sized to exact character counts on the Hub, so loading
# each one whole gives the exact 1:2 ratio — no slicing needed.
#
# Usage:
#   python train.py --config qwen2.5_0.5b_cpt_mix_1to2.yml
# ==============================================================================

base_model: adityabanerjee13/qwen2.5-0.5b-cpt-mix-1to2
model_type: AutoModelForCausalLM
tokenizer_type: AutoTokenizer
trust_remote_code: false

adapter:
load_in_8bit: false
load_in_4bit: false

# multi_eval_plugin splits test_datasets back into per-source eval sets so
# this run logs eval_indic_cpt_mini_val_loss and eval_fineweb_cpt_val_loss
# separately (instead of one merged eval_loss) at every eval step, incl. to
# wandb. Requires this folder on PYTHONPATH — launch via
# `python train.py --config <this file>`.
plugins:
  - multi_eval_plugin.MultiEvalPlugin

datasets:
  - path: adityabanerjee13/indic-sft-mini-train
    type: chat_template
    field_messages: messages
    split: train
  - path: adityabanerjee13/tulu-sft-mini-train
    type: chat_template
    field_messages: messages
    split: train

test_datasets:
  - path: adityabanerjee13/indic-sft-mini-val
    type: chat_template
    field_messages: messages
    split: validation
  - path: adityabanerjee13/tulu-sft-mini-val
    type: chat_template
    field_messages: messages
    split: validation

train_on_inputs: false   

chat_template: tokenizer_default

dataset_prepared_path: ./last_run_prepared_1to2
dataset_num_proc: 1          # single-process tokenize: avoids fork deadlock
val_set_size: 0
output_dir: ./outputs/qwen2.5-0.5b-sft-IT

# --- Sequence packing -----------------------------------------------------
sequence_len: 4096
sample_packing: true
pad_to_sequence_len: true
eval_sample_packing: false

# --- Optimization ----------------------------------------------------------
gradient_accumulation_steps: 8
micro_batch_size: 4
num_epochs: 3
optimizer: adamw_torch_fused
lr_scheduler: cosine
learning_rate: 2e-5
warmup_ratio: 0.03
weight_decay: 0.01
max_grad_norm: 1.0

train_on_inputs: true
group_by_length: false

# --- Precision / memory ---------------------------------------------------
bf16: auto
fp16:
tf32: true
gradient_checkpointing: true
flash_attention: true

# --- Logging / checkpoints ------------------------------------------------
logging_steps: 10
save_strategy: steps
save_steps: 500
save_total_limit: 30
save_only_model: true   # save weights only — no optimizer/scheduler state
                        # (checkpoints ~1/3 the size; can't resume training)
evals_per_epoch: 4

wandb_project: indic-sft
wandb_entity: models-na9841
wandb_name: qwen2.5-0.5b-sft-IT
wandb_log_model: "false"

hub_model_id: adityabanerjee13/qwen2.5-0.5b-sft-IT
hub_strategy: all_checkpoints

special_tokens: