GLM-5.3 is live. Run Z.ai's latest model on Friendli Model APIs. Try it today ➜
haohuiliang
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Run this model inference with full control and performance in your environment.
Model Details
Model Provider
haohuiliang
Model Tree
MidTool/Arctic-MidTool-RL-4B
Input Modalities
Output Modalities
Supported Functionality
Training recipe · Environment and dependencies
692 successful DeepSeek-V4.1-Flash teacher trajectories from 80% Train, with separate demonstration coverage of 357/1,025 clean task–tool-configuration cases (34.83%) and 335/1,025 fault cases (32.68%). Selective decision supervision applies: identified invalid decisions remain in context with their target loss masked, while valid fault encounters and recovery remain learnable. Training uses batch 8, learning rate 3e-6, cosine decay, 8% warmup and three epochs. The adopted checkpoint is epoch 2 / step 174. The initial model is Arctic-MidTool-RL-4B at revision 38b5e742fd20023058cd2cb207358dd033556773.
Training uses two 96 GB GPUs and a 32K context. SFT/RL share 15 validation records for supervised loss diagnostics, without training gradients. Weights are stored as FP32 and evaluated with BF16 compute.
| Measure | Result |
|---|---|
| Clean task success | 28/130 (21.54%) |
| Actual injection coverage | 121/130 (93.08%) |
| Actual-hit task success | 29/121 (23.97%) |
| Base/SFT/RL/teacher common-hit success | 28/116 (24.14%) |
| Four-model common ignored-filter OOD success | 4/20 (20.00%) |
Evaluation fixes tasks, state, verifiers, fault mechanisms and repetition limits. Only valid zero-hit runs under these fixed settings receive supplementation, stopping on the first valid hit even when the task fails. Recovery additionally requires matched clean success. The ignored-filter family is held out of project SFT/RL.
python
import torchfrom transformers import AutoModelForCausalLM, AutoTokenizer repo = "haohuiliang/RobustOps-SFT-4B"revision = "main"tokenizer = AutoTokenizer.from_pretrained(repo, revision=revision)model = AutoModelForCausalLM.from_pretrained( repo, revision=revision, dtype=torch.bfloat16, device_map="auto",)Use the supplied tool-call chat template and tool definitions, with thinking disabled. Running enterprise tasks also requires the RobustOps runtime and adapters.
Research in a stateful enterprise simulator. Aggregate evaluation results are reported in the main README.
Apache-2.0 with upstream attribution. Base model · EnterpriseOps-Gym · veRL.