Modelpublic

laion/tt-x3_kl-kl0

sourceHugging Faceupdated 1mo agoView on Hugging Face

0likes97downloads

Model Card

tt-x3_kl-kl0

RL-finetuned from Qwen/Qwen3-Coder-30B-A3B-Instruct using RLOO with FSDP2 expert-parallel training on the exp_rpt_multifile terminal-bench agentic task suite (terminus-2 Harbor harness). This is the kl_loss_coef=0.0 (no-KL control) arm of the X3 KL-coefficient sweep.

Training configuration

parameter	value
algorithm	RLOO (n=8)
strategy	FSDP2 + expert-parallel (EP=4)
maxgradnorm	1.0
learning_rate	8e-6
eps_clip	low=0.2, high=0.05
loss_reduction	seqmeantokensumnorm_global
TIS	enabled (cap=2.0)
KL loss	disabled (coef=0.0)
batch_size	64 groups x 8 samples
max_steps	80 (reached 38 -- NCCL collective stall)
selected checkpoint	globalstep35

Training Traces

Companion trace dataset: DCAgent/tt-x3_kl-kl0