cemig-nlp-releases/energy-gpt-regulatorio-v3

base_model: ./outputs/qwen3.5-4b-regulacao-distribuicao-mcsynth-1e5
# Stage 2: full SFT refinement on domain data (regulacao + distribuicao).
# Requires stage 1 (qwen3.5-cemig.yml) to have finished and saved weights.

plugins:
  - axolotl.integrations.cut_cross_entropy.CutCrossEntropyPlugin
strict: false

chat_template: qwen3_5
datasets:
  - path: TokenLab/CemigConvoV1.1
    split: regulacao
    type: chat_template
    field_messages: messages
    roles_to_train: [assistant]

  - path: /workspace/pipeline/data/distribuicao/norm_tecnicas_dist.jsonl
    ds_type: json
    type: chat_template
    field_messages: messages
    roles_to_train: [assistant]

val_set_size: 0.05
output_dir: ./outputs/qwen3.5-4b-regulacao-distribuicao-stage2-5e6
dataset_prepared_path: ./prepared_datasets/qwen3.5-4b-stage2

sequence_len: 4096
sample_packing: true

# Freeze vision encoder (same as stage 1)
unfrozen_parameters:
  - model.language_model.*
  - lm_head.*

wandb_project: qwen3.5-4b-regulacao-distribuicao
wandb_entity:
wandb_watch:
wandb_name: qwen3.5-4b-regulacao-distribuicao-stage2
wandb_log_model:

gradient_accumulation_steps: 4
micro_batch_size: 8
num_epochs: 1
optimizer: adamw_torch_fused
lr_scheduler: cosine
learning_rate: 5e-6

bf16: auto
tf32: true

gradient_checkpointing: true
gradient_checkpointing_kwargs:
  use_reentrant: false
resume_from_checkpoint:
logging_steps: 1
attn_implementation: flash_attention_2

warmup_ratio: 0.02
evals_per_epoch: 2
saves_per_epoch: 1
weight_decay: 0.01
special_tokens:

# fsdp_config:
#   fsdp_version: 2
#   offload_params: false
#   cpu_ram_efficient_loading: false
#   auto_wrap_policy: TRANSFORMER_BASED_WRAP
#   transformer_layer_cls_to_wrap: Qwen3_5DecoderLayer
#   state_dict_type: FULL_STATE_DICT
#   sharding_strategy: FULL_SHARD
#   reshard_after_forward: true
#   activation_checkpointing: true