See axolotl config

axolotl version: 0.12.0

# Name wildchat-reconstructed-query-unique-sft_query_generation-qwen3_8b_base

# axolotl train red_team_agent/claude_wildchat/query_gen_stage2_unique.yaml


base_model: Qwen/Qwen3-8B-Base
model_type: AutoModelForCausalLM
tokenizer_type: AutoTokenizer
trust_remote_code: false

# --- Dataset Configuration ---
datasets:
  - path: nate-rahn/wildchat-reconstructed-query-unique
    type: chat_template # Use the chat_template processing strategy
    # --- Custom Template & Role Mapping ---
    chat_template: chatml # Specify we are using a custom jinja template below
    field_messages: messages # Assumes your dataset has a "messages" key with a list of dicts
    message_property_mappings: # Assumes each dict in the list has "role" and "content" keys
      role: role
      content: content
    roles: # Define the roles expected in your dataset for mapping
      user: ["user"] # Map "user" role in data to internal "user"
      assistant: ["assistant"] # Map "assistant" role in data to internal "assistant"
      system: ["system"] # Map "system" role in data to internal "system"
    # --- Training Target ---
    roles_to_train: ["assistant"]
    train_on_eos: turn # Train on the EOS token at the end of each 'user' turn

dataset_prepared_path: /scratch/tmp/wildchat_reconstructed_query_unique_sft/last_run_prepared
dataset_processes: 128

# --- Training Hyperparameters ---
sequence_len: 4096 # Adjust based on your dataset and GPU memory
sample_packing: true # Pack multiple sequences into one example for efficiency
eval_sample_packing: true
pad_to_sequence_len: true # Pad sequences to sequence_len

# Full Parameter Finetuning (No adapter specified)
# adapter: # This is intentionally left blank/removed for full finetuning

# Performance & Precision (H100s excel with bf16)
bf16: true
tf32: true
flash_attention: true # for qwen

# Batching (Adjust based on GPU memory)
# Effective global batch size = micro_batch_size * gradient_accumulation_steps * num_gpus (4)
# Start low for full finetuning, e.g., 1 * 16 * 4 = 64
micro_batch_size: 2
gradient_accumulation_steps: 32
eval_batch_size: 16 # Can often be slightly higher than micro_batch_size

# Optimizer & Scheduler
optimizer: adamw_torch_fused # Good choice for newer GPUs
learning_rate: 1e-5 # Common starting point for full SFT
weight_decay: 0.01
lr_scheduler: cosine # Standard scheduler
warmup_steps: 50
max_grad_norm: 1.0

# Training Duration & Evaluation/Saving
num_epochs: 1 # Train for 1 epoch as requested
val_set_size: 0.01
logging_steps: 1
evals_per_epoch: 20
saves_per_epoch: 2 # Save 2 times per epoch
save_total_limit: 1 # Keep only the last 1 checkpoints

# Memory Saving
# gradient_checkpointing: true # Essential for full finetuning
# gradient_checkpointing_kwargs:
#   use_reentrant: false # Prefer non-reentrant if possible

# --- FSDP Configuration (for 4xH100) ---
fsdp:
  - full_shard
  - auto_wrap
fsdp_config:
  fsdp_offload_params: false # Should not be needed with H100 VRAM
  fsdp_sync_module_states: true # Important for correctness
  fsdp_use_orig_params: false # Recommended for memory saving with FSDP
  fsdp_state_dict_type: SHARDED_STATE_DICT # Options: FULL_STATE_DICT or SHARDED_STATE_DICT (saves disk space)
  fsdp_transformer_layer_cls_to_wrap: 'Qwen3DecoderLayer'
  fsdp_activation_checkpointing: true # Alternative way to enable activation checkpointing for FSDP

# --- Special Tokens ---
# Define based on your custom template's terminators. Qwen already uses <|im_end|>
special_tokens:
  eos_token: "<|im_end|>"

# --- Logging & Saving ---
output_dir: /scratch/out/red-team-agent/runs/wildchat-reconstructed-query-unique-generator-qwen3_8b_base # Local output directory

# W&B Logging
wandb_project: "red-team-agent" # Name your W&B project
wandb_entity: "aqi1048576-mats-program" # IMPORTANT: Replace with your W&B username or team name
wandb_name: "wildchat-reconstructed-query-unique-generator-qwen3_8b_base" # Descriptive run name
# wandb_log_model: "checkpoint" # Log model checkpoints to W&B Artifacts

# Hugging Face Hub Upload
hub_model_id: "nate-rahn/wildchat-reconstructed-query-unique-generator-qwen3_8b_base" # IMPORTANT: Replace with your desired HF repo ID
hub_strategy: "end" # Push checkpoints to the Hub (`"end"` pushes only the final model)
hf_use_auth_token: true # Required for pushing to the Hub (ensure you're logged in)

# --- Misc ---
seed: 42 

wildchat-reconstructed-query-unique-generator-qwen3_8b_base

This model is a fine-tuned version of Qwen/Qwen3-8B-Base on the nate-rahn/wildchat-reconstructed-query-unique dataset. It achieves the following results on the evaluation set:

• Loss: 1.1314
• Memory/max Mem Active(gib): 47.35
• Memory/max Mem Allocated(gib): 46.98
• Memory/device Mem Reserved(gib): 57.16

Model description

More information needed

Intended uses & limitations

More information needed

Training and evaluation data

More information needed

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 1e-05
• train_batch_size: 2
• eval_batch_size: 16
• seed: 42
• distributed_type: multi-GPU
• num_devices: 8
• gradient_accumulation_steps: 32
• total_train_batch_size: 512
• total_eval_batch_size: 128
• optimizer: Use OptimizerNames.ADAMW_TORCH_FUSED with betas=(0.9,0.999) and epsilon=1e-08 and optimizer_args=No additional optimizer arguments
• lr_scheduler_type: cosine
• lr_scheduler_warmup_steps: 50
• training_steps: 106

Training results

Training Loss	Epoch	Step	Validation Loss	Mem Active(gib)	Mem Allocated(gib)	Mem Reserved(gib)
No log	0	0	1.7080	18.69	18.31	24.71
1.6601	0.0561	6	1.6957	47.35	46.98	57.16
1.5452	0.1123	12	1.5717	47.35	46.98	57.16
1.4188	0.1684	18	1.4456	47.35	46.98	57.16
1.3673	0.2246	24	1.3831	47.35	46.98	57.16
1.3088	0.2807	30	1.3335	47.35	46.98	57.16
1.2925	0.3368	36	1.2984	47.35	46.98	57.16
1.2483	0.3930	42	1.2701	47.35	46.98	57.16
1.2152	0.4491	48	1.2444	47.35	46.98	57.16
1.1843	0.5053	54	1.2190	47.35	46.98	57.16
1.1673	0.5614	60	1.1950	47.35	46.98	57.16
1.1662	0.6175	66	1.1751	47.35	46.98	57.16
1.1309	0.6737	72	1.1591	47.35	46.98	57.16
1.1186	0.7298	78	1.1472	47.35	46.98	57.16
1.1299	0.7860	84	1.1393	47.35	46.98	57.16
1.1246	0.8421	90	1.1347	47.35	46.98	57.16
1.1127	0.8982	96	1.1323	47.35	46.98	57.16
1.1011	0.9544	102	1.1314	47.35	46.98	57.16

Framework versions

• Transformers 4.55.0
• Pytorch 2.6.0+cu126
• Datasets 4.0.0
• Tokenizers 0.21.4