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
haohuiliang/RobustOps-SFT-4B
Input Modalities
Output Modalities
Supported Functionality
Training recipe · Environment and dependencies
Initialized from RobustOps-SFT-4B, then trained for 120 updates on Train tasks. Each update uses 16 clean and 16 fault groups with four members per group, learning rate 1e-6 and rollout temperature 1. Fault siblings continue from the same actual post-fault state to learn recovery.
Success, dependency-aware progress and successful-action quality use weights 1 / 0.3 / 0.1, normalized separately following GDPO. The initial model is Arctic-MidTool-RL-4B.
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 BF16, converted from the adopted FP32 export.
The reference scores below come from the original FP32 export evaluated in BF16. Every converted tensor matches its source BF16 cast; the serialized derivative has no separate benchmark.
| Measure | Result |
|---|---|
| Clean task success | 40/130 (30.77%) |
| Actual injection coverage | 125/130 (96.15%) |
| Actual-hit task success | 42/125 (33.60%) |
| Base/SFT/RL/teacher common-hit success | 39/116 (33.62%) |
| Four-model common ignored-filter OOD success | 7/20 (35.00%) |
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.
RL uses 30.63% more completion tokens and 30.86% more tool calls than SFT. Four-model matched recovery is RL 9/10 versus SFT 10/10.
python
import torchfrom transformers import AutoModelForCausalLM, AutoTokenizer repo = "haohuiliang/RobustOps-RL-4B-BF16"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.
import torchfrom transformers import AutoModelForCausalLM, AutoTokenizer repo = "haohuiliang/RobustOps-RL-4B-BF16"revision = "main"tokenizer = AutoTokenizer.from_pretrained(repo, revision=revision)model = AutoModelForCausalLM.from_pretrained( repo, revision=revision, dtype=torch.bfloat16, device_map="auto",)