4b-strat-300-LH-27b-z-e2-iter3-epoch3

<!-- jh-workflow-training-begin -->

Training provenance

Dataset: Stage-org/4b-strat-300-LH-27b-z-e2-iter3
Attempt: 1

Training command

bash /NHNHOME/shkim/stage_tmp/scripts/train_learner.sh /NHNHOME/shkim/jh-workflow/runtime/runs/4b-strat-300-LH-27b-z-e2-iter3/attempt-0001/train.effective.toml Stage-org/4b-strat-300-LH-27b-z-e2-iter3 /NHNHOME/shkim/stage_tmp/outputs/jh-workflow/4b-strat-300-LH-27b-z-e2-iter3/attempt-0001 --experiment-name 4b-strat-300-LH-27b-z-e2-iter3 --execute

Training config

"schema_version" = "stage.config.v7"
"dataset" = { "type" = "new_task", "path" = "", "split" = "train", "token_env" = "HF_TOKEN" }
"learner" = { "method" = "rl", "model" = "Stage-org/4b-strat-300-LH-27b-z-iter2-epoch3-agent-rl-epoch2-low_lr", "learner_steps" = 10000, "learner_epoch" = 3, "batch_size" = 128, "seq_len" = 300000, "wandb" = { "enabled" = true, "project" = "stage-learner-junhee", "name" = "learner training", "group" = "stage-experiment", "tags" = ["learner"] }, "checkpoint" = { "output_dir" = "checkpoint-store", "interval" = 1000, "interval_unit" = "epoch", "keep_last" = 1 }, "rl" = { "group_size" = 8, "weight_broadcast_type" = "filesystem", "gpus_per_node" = 2, "num_infer_gpus" = 1, "num_train_gpus" = 1, "environment_id" = "", "generation" = { "temperature" = 0.9, "max_tokens" = 4096, "top_p" = 1.0, "enable_thinking" = true }, "open_ended_judge" = { "mean_score" = false, "max_retries" = 3, "retry_backoff_seconds" = 1.0, "max_in_flight" = 32, "endpoint" = { "model" = "gpt-5.6-luna", "base_url_env" = "JUDGE_BASE_URL", "api_key_env" = "JUDGE_API_KEY" }, "generation" = { "temperature" = 1.0, "max_tokens" = 4096, "top_p" = 1.0, "reasoning_effort" = "medium" } } } }
"loop" = { "seed" = 7 }
"prime_rl" = { "rl" = { "trainer" = { "ckpt" = { "weights_only" = true }, "model" = { "attn" = "flash_attention_2" }, "optim" = { "type" = "adamw", "lr" = 1e-06, "weight_decay" = 0.0, "max_norm" = 1.0, "betas1" = 0.9, "betas2" = 0.99 }, "loss" = { "type" = "default", "dppo_mask_low" = 0.2, "dppo_mask_high" = 0.28, "adv_tau" = 1.0, "kl_tau" = 0.001 } }, "inference" = { "gpu_memory_utilization" = 0.9, "model" = { "max_model_len" = 65536 }, "vllm_extra" = { "language_model_only" = true, "reasoning_parser" = "qwen3", "tool_call_parser" = "qwen3_coder" }, "server" = { "port" = 7000 } }, "orchestrator" = { "max_inflight_rollouts" = 256, "max_off_policy_steps" = 8, "model" = { "client" = { "base_url" = ["http://localhost:7000/v1"] } } } } }

<!-- jh-workflow-training-end -->