metadata
dataset_info:
- config_name: hyperparameters__rl
features:
- name: stage_name
dtype: string
- name: stage_number
dtype: int64
- name: stage_type
dtype: string
- name: model_repo_id
dtype: string
- name: base_model
dtype: string
- name: timestamp
dtype: string
- name: verl_parameter_config
struct:
- name: actor_rollout_ref.actor.fsdp_config.forward_prefetch
dtype: bool
- name: actor_rollout_ref.actor.optim.lr
dtype: float64
- name: actor_rollout_ref.actor.ppo_micro_batch_size_per_gpu
dtype: int64
- name: actor_rollout_ref.actor.ppo_mini_batch_size
dtype: int64
- name: actor_rollout_ref.actor.strategy
dtype: string
- name: actor_rollout_ref.model.enable_activation_offload
dtype: bool
- name: actor_rollout_ref.model.enable_gradient_checkpointing
dtype: bool
- name: actor_rollout_ref.model.path
dtype: string
- name: actor_rollout_ref.model.trust_remote_code
dtype: bool
- name: actor_rollout_ref.model.use_remove_padding
dtype: bool
- name: actor_rollout_ref.ref.fsdp_config.forward_prefetch
dtype: bool
- name: actor_rollout_ref.ref.log_prob_micro_batch_size_per_gpu
dtype: int64
- name: actor_rollout_ref.rollout.dtype
dtype: string
- name: actor_rollout_ref.rollout.gpu_memory_utilization
dtype: float64
- name: actor_rollout_ref.rollout.log_prob_micro_batch_size_per_gpu
dtype: int64
- name: actor_rollout_ref.rollout.max_num_batched_tokens
dtype: int64
- name: actor_rollout_ref.rollout.max_num_seqs
dtype: int64
- name: actor_rollout_ref.rollout.n
dtype: int64
- name: actor_rollout_ref.rollout.tensor_model_parallel_size
dtype: int64
- name: algorithm.adv_estimator
dtype: string
- name: algorithm.kl_ctrl.kl_coef
dtype: float64
- name: critic.model.path
dtype: string
- name: critic.model.trust_remote_code
dtype: bool
- name: critic.optim.lr
dtype: float64
- name: critic.ppo_micro_batch_size_per_gpu
dtype: int64
- name: custom_reward_function.name
dtype: string
- name: custom_reward_function.path
dtype: string
- name: custom_reward_function.reward_kwargs.complex_format_reward_weight
dtype: float64
- name: >-
custom_reward_function.reward_kwargs.final_answer_in_samples_reward_weight
dtype: float64
- name: >-
custom_reward_function.reward_kwargs.reflection_correctness_reward_weight
dtype: float64
- name: custom_reward_function.reward_kwargs.response_or_sample
dtype: string
- name: custom_reward_function.reward_kwargs.reward_max
dtype: float64
- name: custom_reward_function.reward_kwargs.reward_min
dtype: float64
- name: >-
custom_reward_function.reward_kwargs.sample_correctness_reward_weight
dtype: float64
- name: custom_reward_function.reward_kwargs.sample_count_penalty_weight
dtype: float64
- name: custom_reward_function.reward_kwargs.similarity_penalty_weight
dtype: float64
- name: custom_reward_function.reward_kwargs.simple_format_reward_weight
dtype: float64
- name: custom_reward_function.reward_kwargs.transition_penalty_weight
dtype: float64
- name: >-
custom_reward_function.reward_kwargs.verdict_correctness_reward_weight
dtype: float64
- name: data.max_prompt_length
dtype: int64
- name: data.max_response_length
dtype: int64
- name: data.train_batch_size
dtype: int64
- name: data.train_files
dtype: string
- name: data.val_files
dtype: string
- name: hydra.job.chdir
dtype: bool
- name: hydra.output_subdir
dtype: string
- name: hydra.run.dir
dtype: string
- name: reward_model.launch_reward_fn_async
dtype: bool
- name: reward_model.model.fsdp_config.forward_prefetch
dtype: bool
- name: reward_model.reward_manager
dtype: string
- name: trainer.default_local_dir
dtype: string
- name: trainer.experiment_name
dtype: string
- name: trainer.logger
dtype: string
- name: trainer.n_gpus_per_node
dtype: int64
- name: trainer.nnodes
dtype: int64
- name: trainer.project_name
dtype: string
- name: trainer.resume_mode
dtype: string
- name: trainer.save_freq
dtype: int64
- name: trainer.test_freq
dtype: int64
- name: trainer.total_epochs
dtype: int64
- name: trainer.val_before_train
dtype: bool
splits:
- name: train
num_bytes: 1382
num_examples: 1
download_size: 41655
dataset_size: 1382
- config_name: logs__verl_rl
features:
- name: timestamp
dtype: string
- name: end_timestamp
dtype: string
- name: stage_name
dtype: string
- name: stage_number
dtype: int64
- name: level
dtype: string
- name: message
dtype: string
- name: stdout_content
dtype: string
- name: stderr_content
dtype: string
- name: experiment_name
dtype: string
- name: elapsed_time_seconds
dtype: float64
- name: stage_complete
dtype: bool
splits:
- name: train
num_bytes: 3755076
num_examples: 2
download_size: 624034
dataset_size: 3755076
- config_name: metadata
features:
- name: experiment_name
dtype: string
- name: start_time
dtype: string
- name: description
dtype: string
- name: base_org
dtype: string
- name: stage_number
dtype: string
- name: stage_type
dtype: string
- name: status
dtype: string
splits:
- name: train
num_bytes: 2644
num_examples: 7
download_size: 4615
dataset_size: 2644
- config_name: training_data__rl_metadata
features:
- name: stage_name
dtype: string
- name: stage_number
dtype: int64
- name: timestamp
dtype: string
- name: original_dataset_id
dtype: string
- name: dataset_type
dtype: string
- name: rl_training_splits
list: string
- name: rl_validation_splits
list: string
- name: rl_configs
list: string
- name: usage
dtype: string
splits:
- name: train
num_bytes: 202
num_examples: 1
download_size: 4991
dataset_size: 202
configs:
- config_name: hyperparameters__rl
data_files:
- split: train
path: hyperparameters__rl/train-*
- config_name: logs__verl_rl
data_files:
- split: train
path: logs__verl_rl/train-*
- config_name: metadata
data_files:
- split: train
path: metadata/train-*
- config_name: training_data__rl_metadata
data_files:
- split: train
path: training_data__rl_metadata/train-*
Experiment Tracker: 1123_newmodels__olmo7b_sft_r1_ct3arg
Experiment Description: Experiment: 1123_newmodels__olmo7b_sft_r1_ct3arg
Start Time: 2025-11-26T13:37:11.086300
Tracker Dataset: SkillFactory/D-ExpTracker__1123_newmodels__olmo7b_sft_r1_ct3arg__v1
Stages Completed
Total stages: 1
Models Created
Dataset Configurations
This tracker dataset contains the following configurations with immediate upload as stages complete:
Training Data (Complete Datasets)
Hyperparameters (Complete Configurations)
Logs (Stage-Specific)
Evaluation Results (Complete with Annotations)
Metadata
- experiment_metadata: Timeline and stage information
Usage
Load specific configurations with:
from datasets import load_dataset
# Load experiment metadata
metadata = load_dataset('SkillFactory/D-ExpTracker__1123_newmodels__olmo7b_sft_r1_ct3arg__v1', 'experiment_metadata')
# Load complete training datasets
sft_data = load_dataset('SkillFactory/D-ExpTracker__1123_newmodels__olmo7b_sft_r1_ct3arg__v1', 'training_data__sft')
sft_metadata = load_dataset('SkillFactory/D-ExpTracker__1123_newmodels__olmo7b_sft_r1_ct3arg__v1', 'training_data__sft_metadata')
# Load complete configurations
sft_hyperparams = load_dataset('SkillFactory/D-ExpTracker__1123_newmodels__olmo7b_sft_r1_ct3arg__v1', 'hyperparameters__sft')
rl_hyperparams = load_dataset('SkillFactory/D-ExpTracker__1123_newmodels__olmo7b_sft_r1_ct3arg__v1', 'hyperparameters__rl')
# Load stage-specific logs
sft_logs = load_dataset('SkillFactory/D-ExpTracker__1123_newmodels__olmo7b_sft_r1_ct3arg__v1', 'logs__sft')
rl_logs = load_dataset('SkillFactory/D-ExpTracker__1123_newmodels__olmo7b_sft_r1_ct3arg__v1', 'logs__rl')
# Load evaluation results with annotations
sft_eval_results = load_dataset('SkillFactory/D-ExpTracker__1123_newmodels__olmo7b_sft_r1_ct3arg__v1', 'evals_eval_sft')
rl_eval_results = load_dataset('SkillFactory/D-ExpTracker__1123_newmodels__olmo7b_sft_r1_ct3arg__v1', 'evals_eval_rl')
Models
Registry
All models from this experiment are automatically registered in the SkillFactory Model Registry with:
- Complete training configuration (hyperparameters, datasets, methods)
- Experiment lineage (links back to this tracker dataset)
- Stage-specific metadata (SFT vs RL training details)
- Structured input data references (training datasets and configurations)
Registry entries follow the naming pattern: Model - 1123_newmodels__olmo7b_sft_r1_ct3arg - {stage_name} - {SFT/RL}
Generated by SkillFactory Experiment Management System All artifacts uploaded immediately as stages complete with perfect data provenance