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.
Run this model inference with full control and performance in your environment.
Model Details
Model Provider
closestfriend
Model Tree
Qwen/Qwen2.5-3B-Instruct
Input Modalities
Output Modalities
Supported Functionality
This adapter is the canonical Qwen 2.5 3B artifact used by the Brie paper and
evaluation pipeline. closestfriend/brie-v2-qwen2.5-3b is a later full-weight
distribution produced by merging this adapter into the base model; it is not a
separate result.
The comprehensive evaluation contains 57 judgments over 25 unique prompts (none of which appear in the training data); configuration blocks repeat prompts. Judges received blind, position-randomized A/B comparisons of the same base and Brie responses. Primary rates retain ties and assigned prompts in the denominator.
| Panel | Brie preference | Construction |
|---|---|---|
| 2025 primary | 88.9% | Mean of three full n=57 passes: Claude Opus 4, GPT-4o, and Gemini 2.5 Flash Lite |
| 2025 expanded sensitivity | 90.5% | Adds Claude 3.5 Sonnet (n=42) to the three full passes; prompt coverage is unequal |
| 2026 primary | 78.9% | Four-model mean, n=57 per model: Claude Haiku 4.5, Claude Sonnet 4.5, GPT-5, and Gemini 3 Pro |
The historical 91.2% result is the original split-source Anthropic pool: Claude 3.5 Sonnet judged 42 prompts and Claude Opus 4 judged the other 15, yielding 52/57 Brie preferences. It is not a uniform multi-judge mean. A separate GPT-4o/Gemini verdict-agreement calculation also happens to equal 91.2% (52/57).
The small n=15 out-of-domain control provides no detected broad regression under the tested panels, but is too small to establish absence of capability loss.
Full pass identities, counts, hashes, and reporting decisions are documented
in the paper repository's docs/QWEN_3B_2025_JUDGE_PROVENANCE.md and generated
claim ledger.
Qwen/Qwen2.5-3B-Instructq_proj, v_projpython
from transformers import AutoModelForCausalLM, AutoTokenizerfrom peft import PeftModel base_id = "Qwen/Qwen2.5-3B-Instruct"adapter_id = "closestfriend/brie-v2-3b" 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 research artifact. The corpus reflects one author's English-language philosophical and creative practice; it is not a general factual-authority or safety-tuned model.
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/Qwen2.5-3B-Instruct"adapter_id = "closestfriend/brie-v2-3b" 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}}