Downloads · 30 days
0
Zhongzhu/tunekv-taubench-grpo
tunekv-taubench-grpo is a machine learning model from Zhongzhu. Use it for the machine learning task on the model card, and read the license before you ship it in a product.
Merged TP1 prefixkv serving artifacts for the three τ-bench outcome-RL (GRPO clip 0.2 + 0.1 reverse KL) arms. Each subdirectory is a complete serving artifact (tensors + READY marker + composition/proof files).
Downloads · 30 days
0
Access
Public
Updated Sep 23, 2026
Repo size
1.3 GB
Likes
0
Public
Click a slice to open those files.
.safetensors1.3 GB · 100%
From the Hugging Face model README
Merged TP1 prefix_kv serving artifacts for the three τ-bench outcome-RL
(GRPO clip 0.2 + 0.1 reverse KL) arms. Each subdirectory is a complete
serving artifact (tensors + READY marker + composition/proof files).
| Subdirectory | Arm | Base → tuned (same stack) | Files |
|---|---|---|---|
airline30_prefix_kv_tp1/ | airline · Qwen3-Coder-30B | full-50 0.46 → 0.42 | 49 |
retail30_prefix_kv_tp1/ | retail · Qwen3-Coder-30B | test-115 0.4087 → 0.4174 | 49 |
retail27_prefix_kv_tp1/ | retail · Qwen3.8-27B | test-115 0.5565 → 0.5913 | 67 (16 FA + 48 GDN) |
Training rows, rollouts, eval results, configs and logs:
Zhongzhu/tunekv-taubench-data
(grpo_airline30/, grpo_retail30/, grpo_retail27/).
Serve with the TuneKV vLLM connector (tunekv_vllm.hybrid for 27B,
dense connector for 30B), n_inject = 4224 (airline) / 4096 (retail-30B) /
4352 (retail-27B). See each data subdirectory's META.json for the full
recipe (packing windows, step counts, wandb runs, paired flips).