Downloads · 30 days
14
41% of all-time downloads
Stage-jh-monitor/total-32-sampling1-jh
total-32-sampling1-jh is a machine learning model from Stage-jh-monitor. Use it for the machine learning task on the model card, and read the license before you ship it in a product.
Dataset: Stage-org/total-32-sampling1-jh Attempt: 1
Downloads · 30 days
14
41% of all-time downloads
All-time downloads
34
Public
Parameters
4.5B
9.1 GB on disk
Likes
0
Public
Click a slice to open those files.
.safetensors9.1 GB · 100%
From the Hugging Face model README
Dataset: Stage-org/total-32-sampling1-jh
Attempt: 1
bash /NHNHOME/shkim/stage_tmp/scripts/train_learner.sh /NHNHOME/shkim/jh-workflow/configs/learner_rl_b200_jh_workflow.toml Stage-org/total-32-sampling1-jh /NHNHOME/shkim/stage_tmp/outputs/jh-workflow/total-32-sampling1-jh/attempt-0001 --experiment-name total-32-sampling1-jh --execute
schema_version = "stage.config.v7"
[dataset]
type = "new_task"
path = ""
split = "train"
token_env = "HF_TOKEN"
[learner]
method = "rl"
model = "Qwen/Qwen3.5-4B"
learner_steps = 10000
learner_epoch = 6
batch_size = 128
[learner.wandb]
enabled = true
project = "stage-learner-junhee"
name = "learner training"
group = "stage-experiment"
tags = ["learner"]
[learner.checkpoint]
output_dir = "checkpoint-store"
interval = 1000
interval_unit = "epoch"
keep_last = 1
[learner.rl]
group_size = 16
weight_broadcast_type = "filesystem"
gpus_per_node = 2
num_infer_gpus = 1
num_train_gpus = 1
# only for method="agent_rl"
environment_id = ""
[learner.rl.generation]
temperature = 0.9
max_tokens = 4096
top_p = 1.0
enable_thinking = true
[learner.rl.open_ended_judge]
max_retries = 3
retry_backoff_seconds = 1.0
max_in_flight = 16
[learner.rl.open_ended_judge.endpoint]
model = "gpt-5.6-luna"
base_url_env = "JUDGE_BASE_URL"
api_key_env = "JUDGE_API_KEY"
[learner.rl.open_ended_judge.generation]
temperature = 1.0
max_tokens = 4096
top_p = 1.0
reasoning_effort = "medium"
[loop]
# Controls the deterministic ordering of generated learner tasks.
seed = 7
[prime_rl.rl.trainer.ckpt]
weights_only = true
[prime_rl.rl.inference]
gpu_memory_utilization = 0.9
[prime_rl.rl.inference.model]
max_model_len = 65536
[prime_rl.rl.inference.vllm_extra]
language_model_only = true
reasoning_parser = "qwen3"
tool_call_parser = "qwen3_coder"
[prime_rl.rl.trainer.model]
attn = "flash_attention_2"
dp_replicate = 1
[prime_rl.rl.orchestrator]
max_inflight_rollouts = 256
max_off_policy_steps = 8
[prime_rl.rl.trainer.optim]
type = "adamw"
lr = 1e-6
weight_decay = 0.00
max_norm = 0.5
betas1 = 0.9
betas2 = 0.995
[prime_rl.rl.trainer.loss]
type = "default"
dppo_mask_low = 0.2
dppo_mask_high = 0.28
adv_tau = 1.0
kl_tau = 1e-3
[prime_rl.rl.trainer.scheduler]
type = "constant"
<!-- jh-workflow-training-end -->