Dataset Viewer
stage_name
string | stage_number
int64 | stage_type
string | model_repo_id
string | base_model
string | timestamp
string | verl_parameter_config
dict |
|---|---|---|---|---|---|---|
rl
| 1 |
verl_rl_training
|
TAUR-dev/M-SBON_advanced_grpo_rewards-grpo_adv_rwds-rl
|
TAUR-dev/M-skills_in_rl_v2__1e6_all_tasks_sft-sft
|
2025-08-20T18:01:29.110260
|
{
"actor_rollout_ref.actor.fsdp_config.forward_prefetch": true,
"actor_rollout_ref.actor.optim.lr": 0.000005,
"actor_rollout_ref.actor.ppo_micro_batch_size_per_gpu": 8,
"actor_rollout_ref.actor.ppo_mini_batch_size": 64,
"actor_rollout_ref.actor.strategy": "fsdp2",
"actor_rollout_ref.model.enable_activation_offload": true,
"actor_rollout_ref.model.enable_gradient_checkpointing": true,
"actor_rollout_ref.model.path": "/scratch/10416/zaynesprague/skill_inject_outputs/sf_experiments/skills_in_rl/grpo_lexical_sim/all_tasks/SBON_advanced_grpo_rewards/verl/prefetched_models/TAUR_dev__M_skills_in_rl_v2__1e6_all_tasks_sft_sft",
"actor_rollout_ref.model.trust_remote_code": true,
"actor_rollout_ref.model.use_remove_padding": true,
"actor_rollout_ref.ref.fsdp_config.forward_prefetch": true,
"actor_rollout_ref.ref.log_prob_micro_batch_size_per_gpu": 16,
"actor_rollout_ref.rollout.dtype": "bfloat16",
"actor_rollout_ref.rollout.gpu_memory_utilization": 0.8,
"actor_rollout_ref.rollout.log_prob_micro_batch_size_per_gpu": 16,
"actor_rollout_ref.rollout.n": 8,
"actor_rollout_ref.rollout.tensor_model_parallel_size": 1,
"algorithm.adv_estimator": "grpo",
"algorithm.kl_ctrl.kl_coef": 0.001,
"critic.model.path": "/scratch/10416/zaynesprague/skill_inject_outputs/sf_experiments/skills_in_rl/grpo_lexical_sim/all_tasks/SBON_advanced_grpo_rewards/verl/prefetched_models/TAUR_dev__M_skills_in_rl_v2__1e6_all_tasks_sft_sft",
"critic.model.trust_remote_code": true,
"critic.optim.lr": 0.00001,
"critic.ppo_micro_batch_size_per_gpu": 1,
"custom_reward_function.path": "/scratch/10416/zaynesprague/skill_factory_dir/skill-factory/thirdparty/verl/sf_scripts/skill_factory_rewards.py",
"custom_reward_function.reward_kwargs.format_score_v2_weight": 0,
"custom_reward_function.reward_kwargs.format_score_weight": 0,
"custom_reward_function.reward_kwargs.reward_max": 10,
"custom_reward_function.reward_kwargs.reward_min": -1,
"custom_reward_function.reward_kwargs.sample_correctness_weight": 1,
"custom_reward_function.reward_kwargs.sample_count_penalty_weight": 1,
"custom_reward_function.reward_kwargs.similarity_penalty_weight": 1,
"custom_reward_function.reward_kwargs.transition_penalty_weight": 0,
"data.max_prompt_length": 512,
"data.max_response_length": 4096,
"data.train_batch_size": 256,
"data.train_files": "/scratch/10416/zaynesprague/skill_inject_outputs/sf_experiments/skills_in_rl/grpo_lexical_sim/all_tasks/SBON_advanced_grpo_rewards/verl/data/train.parquet",
"data.val_files": "/scratch/10416/zaynesprague/skill_inject_outputs/sf_experiments/skills_in_rl/grpo_lexical_sim/all_tasks/SBON_advanced_grpo_rewards/verl/data/test.parquet",
"hydra.job.chdir": false,
"hydra.output_subdir": "null",
"hydra.run.dir": "/scratch/10416/zaynesprague/skill_injection_outputs/sf/grpo_training/exp1/hydra/grpo",
"reward_model.model.fsdp_config.forward_prefetch": true,
"trainer.default_local_dir": "/scratch/10416/zaynesprague/skill_inject_outputs/sf_experiments/skills_in_rl/grpo_lexical_sim/all_tasks/SBON_advanced_grpo_rewards/verl/checkpoints",
"trainer.experiment_name": "SBON_advanced_grpo_rewards-grpo_adv_rwds_rl",
"trainer.logger": "[console,wandb]",
"trainer.n_gpus_per_node": 1,
"trainer.nnodes": 4,
"trainer.project_name": "rl_skills__8_13_25",
"trainer.resume_mode": "disable",
"trainer.save_freq": 10,
"trainer.test_freq": 10,
"trainer.total_epochs": 30,
"trainer.val_before_train": false
}
|
Experiment Tracker: SBON_advanced_grpo_rewards-grpo_adv_rwds
Experiment Description: GRPO experiment with lexical similarity penalty for Skill Factory workflows.
Start Time: 2025-08-20T17:50:39.752037
Tracker Dataset: TAUR-dev/D-ExpTracker__SBON_advanced_grpo_rewards-grpo_adv_rwds__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('TAUR-dev/D-ExpTracker__SBON_advanced_grpo_rewards-grpo_adv_rwds__v1', 'experiment_metadata')
# Load complete training datasets
sft_data = load_dataset('TAUR-dev/D-ExpTracker__SBON_advanced_grpo_rewards-grpo_adv_rwds__v1', 'training_data__sft')
sft_metadata = load_dataset('TAUR-dev/D-ExpTracker__SBON_advanced_grpo_rewards-grpo_adv_rwds__v1', 'training_data__sft_metadata')
# Load complete configurations
sft_hyperparams = load_dataset('TAUR-dev/D-ExpTracker__SBON_advanced_grpo_rewards-grpo_adv_rwds__v1', 'hyperparameters__sft')
rl_hyperparams = load_dataset('TAUR-dev/D-ExpTracker__SBON_advanced_grpo_rewards-grpo_adv_rwds__v1', 'hyperparameters__rl')
# Load stage-specific logs
sft_logs = load_dataset('TAUR-dev/D-ExpTracker__SBON_advanced_grpo_rewards-grpo_adv_rwds__v1', 'logs__sft')
rl_logs = load_dataset('TAUR-dev/D-ExpTracker__SBON_advanced_grpo_rewards-grpo_adv_rwds__v1', 'logs__rl')
# Load evaluation results with annotations
sft_eval_results = load_dataset('TAUR-dev/D-ExpTracker__SBON_advanced_grpo_rewards-grpo_adv_rwds__v1', 'evals_eval_sft')
rl_eval_results = load_dataset('TAUR-dev/D-ExpTracker__SBON_advanced_grpo_rewards-grpo_adv_rwds__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 - SBON_advanced_grpo_rewards-grpo_adv_rwds - {stage_name} - {SFT/RL}
Generated by SkillFactory Experiment Management System All artifacts uploaded immediately as stages complete with perfect data provenance
- Downloads last month
- 21