GLM-5.2 is live. #1 throughput on OpenRouter, pay-per-token on FriendliAI. Try it today ➜
ThreePoints3
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
ThreePoints3
Model Tree
Qwen/Qwen3-8B
Input Modalities
Output Modalities
Supported Functionality
CPKeys normalizes a user request into a structured specification, generates multiple candidate seeds, and filters them with format checks, quality judging, and CPRet similarity signals. The actor was first supervised-fine-tuned and then updated with an originality-aware RLOO-style objective.
| Field | Value |
|---|---|
| Base model | Qwen/Qwen3-8B |
| Adapter type | LoRA / PEFT |
| Task type | Causal language modeling |
| Rank | 16 |
| Alpha | 32 |
| Dropout | 0.05 |
| Target modules | q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj |
| Checkpoint | RLOO +167 |
The tokenizer and chat template saved with the training checkpoint are included for reproducibility.
python
from transformers import AutoModelForCausalLM, AutoTokenizerfrom peft import PeftModelbase_model_id = "Qwen/Qwen3-8B"adapter_id = "ThreePoints3/CPKeys-Qwen3-8B-RLOO-167"tokenizer = AutoTokenizer.from_pretrained(adapter_id)base_model = AutoModelForCausalLM.from_pretrained(base_model_id,torch_dtype="auto",device_map="auto",)model = PeftModel.from_pretrained(base_model, adapter_id)model.eval()
For the complete normalization, generation, verification, and review workflow, use the scripts and prompts in the GitHub repository rather than calling this adapter without the CPKeys Stage 2 prompt.
The project report compares the raw Qwen3-8B actor, an SFT checkpoint, and RLOO checkpoints on fixed keyword windows and mixed raw inputs. RLOO checkpoints generally increased the density of candidates entering the automatic review queue, but the trend was not monotonic:
See the project report for tables, controls, and limitations.
This adapter is intended for:
Generated outputs are not guaranteed to be correct, original, solvable, or contest-ready. They require similarity checking, technical validation, and human review.
<thinking> / <output> format must be handled consistently by the serving stack.The repository is currently marked other while the project, training-data, and derived-model licensing review is completed. No additional permission is granted by this model card. The Qwen3 base model and CPRetriever dependency retain their respective upstream licenses.