Sweetflips/qwen3-coder-30b-agent-v1

base_model: Qwen/Qwen3-Coder-30B-A3B-Instruct
model_type: AutoModelForCausalLM
tokenizer_type: AutoTokenizer
trust_remote_code: true

load_in_8bit: false
load_in_4bit: false

adapter: lora
lora_r: 128
lora_alpha: 256
lora_dropout: 0.05
lora_target_linear: true
lora_target_modules:
  - q_proj
  - k_proj
  - v_proj
  - o_proj
  - gate_proj
  - up_proj
  - down_proj

datasets:
  - path: /root/sweetflipstraining/data/sweetflippy_train.jsonl
    type: chat_template
    chat_template: chatml
    field_messages: conversations

val_set_size: 0.0
dataset_prepared_path: null
dataset_num_proc: 64
dataloader_num_workers: 8

output_dir: /root/sweetflipstraining/outputs/agent-sft-max-speed
hub_model_id: Sweetflips/qwen3-coder-30b-agent-v1

# REDUCED for memory
sequence_len: 8192
sample_packing: true
pad_to_sequence_len: true

# REDUCED batch size
micro_batch_size: 2
gradient_accumulation_steps: 4

num_epochs: 1
learning_rate: 1e-4
lr_scheduler: cosine
warmup_ratio: 0.05

optimizer: adamw_torch_fused
weight_decay: 0.01
max_grad_norm: 1.0
lr_groups: []

bf16: true
tf32: true

# ENABLE gradient checkpointing to save memory
gradient_checkpointing: true

flash_attention: false
sdp_attention: true

deepspeed: configs/deepspeed_configs/zero3_no_offload.json

save_strategy: steps
save_steps: 100
save_total_limit: 10
eval_strategy: "no"

wandb_mode: disabled
seed: 42
chat_template: chatml
special_tokens:
  pad_token: "<|endoftext|>"

auto_resume_from_checkpoints: false