GLM-5.3 is live. Run Z.ai's latest model on Friendli Model APIs. Try it today ➜
closestfriend
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Model Details
Model Provider
closestfriend
Model Tree
Qwen/Qwen2.5-0.5B-Instruct
Input Modalities
Output Modalities
Supported Functionality
The comprehensive evaluation contains 57 blind, position-randomized A/B judgments over 25 unique prompts (none in the training data). Primary rates retain ties and all assigned prompts.
| Result | Brie preference | Construction |
|---|---|---|
| 2025 primary | 67.8% | Mean of three full n=57 passes: Claude Opus 4, GPT-4o, and Gemini 2.5 Flash Lite |
| Expanded sensitivity | 69.9% | Adds Claude 3.5 Sonnet (n=42); prompt coverage is unequal |
| Historical split-source result | 71.9% | Original disjoint Claude pool, 41/57 |
One record in the historical split source was stored as unknown because its
judgment said Winner: Neither rather than the requested A/B/Tie token. It is
a semantic tie and remains in the denominator. Deleting it produces the older
73.2% parser-exclusion figure; averaging that parser-exclusion source with GPT
and Gemini produces 77.0%. Neither is the preferred headline.
The three full-pass inputs to the 67.8% result are Claude Opus 4 at 26/57, GPT-4o at 43/57, and Gemini 2.5 Flash Lite at 47/57. The separate n=13 in-domain and n=15 out-of-domain suites are not merged into this comprehensive panel.
Complete pass identities, hashes, and reporting decisions are documented in
docs/QWEN_0_5B_JUDGE_PROVENANCE.md in the paper repository.
Qwen/Qwen2.5-0.5B-Instructadapter_config.json is authoritative)q_proj, v_projpython
from transformers import AutoModelForCausalLM, AutoTokenizerfrom peft import PeftModel base_id = "Qwen/Qwen2.5-0.5B-Instruct"adapter_id = "closestfriend/brie-qwen2.5-0.5b" tokenizer = AutoTokenizer.from_pretrained(base_id)base = AutoModelForCausalLM.from_pretrained(base_id, device_map="auto")model = PeftModel.from_pretrained(base, adapter_id)This is a small research model with substantial capacity limits. Its training corpus reflects one author's English-language philosophical and creative practice; verify factual or high-stakes output independently.
bibtex
@misc{karman_brie_domain_adaptation, author = {Karman, Hunter N.}, title = {Human-Curated Data Authoring with LLMs: A Small-Data Approach to Domain Adaptation}, doi = {10.5281/zenodo.17657737}, url = {https://doi.org/10.5281/zenodo.17657737}}from transformers import AutoModelForCausalLM, AutoTokenizerfrom peft import PeftModel base_id = "Qwen/Qwen2.5-0.5B-Instruct"adapter_id = "closestfriend/brie-qwen2.5-0.5b" tokenizer = AutoTokenizer.from_pretrained(base_id)base = AutoModelForCausalLM.from_pretrained(base_id, device_map="auto")model = PeftModel.from_pretrained(base, adapter_id)@misc{karman_brie_domain_adaptation, author = {Karman, Hunter N.}, title = {Human-Curated Data Authoring with LLMs: A Small-Data Approach to Domain Adaptation}, doi = {10.5281/zenodo.17657737}, url = {https://doi.org/10.5281/zenodo.17657737}}