GLM-5.3 is live. Run Z.ai's latest model on Friendli Model APIs. Try it today ➜
Alepach
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Model Details
Model Provider
Alepach
Model Tree
Qwen/Qwen2.5-7B-Instruct
Input Modalities
Output Modalities
Supported Functionality
The model was trained on 10,000 standard Step-DPO preference pairs sampled
deterministically with seed 42. Each pair contains a mathematical problem, a
reasoning prefix ending at Step N:, a chosen replacement step from a
continuation that reached the reference answer, and the original erroneous
step as the rejected response.
The trainer-ready data is available in
Alepach/fixingDPO as
training/standard_step_dpo_10k_seed42.json.
Qwen/Qwen2.5-7B-Instruct5e-70.40.1Training completed in 36,213 seconds with a final aggregate training loss of 0.2444.