Modelpublic

laion/tt-x5_gradnorm-gn0p9

sourceHugging Faceupdated 1mo agoView on Hugging Face

0likes88downloads

Model Card

tt-x5_gradnorm-gn0p9

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).

Training configuration

parameter	value
algorithm	RLOO (n=8)
strategy	FSDP2 + expert-parallel (EP=4)
maxgradnorm	0.9
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 × 8 samples
max_steps	80 (reached 66 — wall-time limited)
selected checkpoint	globalstep65

Training results

metric	value (step 65)
reward (mean)	0.197
pass@8	0.406
policy_entropy	0.492
ppoclipratio	0.011

Training Traces

Companion trace dataset: DCAgent/tt-x5_gradnorm-gn0p9