GLM-5.2 is live. #1 throughput on OpenRouter, pay-per-token on FriendliAI. Try it today ➜
Subject-Emu-5259
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
Subject-Emu-5259
Model Tree
HuggingFaceTB/SmolLM2-360M-Instruct
Input Modalities
Output Modalities
Supported Functionality
HuggingFaceTB/SmolLM2-360M-Instruct (360M)checkpoints/v2_model)data/train_dpo_v16_combined.jsonl)D17 is a DPO continuation of v16 — an alignment refinement, not new knowledge
injection. Across the 679-pair distribution it sharpened preferences toward
correct code, safe refusals, identity-accurate responses, sound math/reasoning,
clean code style, concise answers, and proper tool usage, while down-ranking
rejected patterns. Chosen-response rewards rose (0.54 → 0.69) and rejected
rewards fell (−0.19 → −0.26) with a healthy ~0.9 margin and stable entropy —
textbook successful DPO with no sign of reward hacking or mode collapse. No
held-out eval set was configured, so generalization is inferred from training
reward signals only.
python
from peft import PeftModelfrom transformers import AutoModelForCausalLM, AutoTokenizerbase = AutoModelForCausalLM.from_pretrained("HuggingFaceTB/SmolLM2-360M-Instruct")model = PeftModel.from_pretrained(base, "Subject-Emu-5259/NeuralAI")tokenizer = AutoTokenizer.from_pretrained("Subject-Emu-5259/NeuralAI")