CPKeys Qwen3-8B RLOO-167
CPKeys Qwen3-8B RLOO-167 is a PEFT LoRA adapter for generating competitive-programming problem seeds from structured keys. It is the RLOO +167 research checkpoint used by the CPKeys pipeline.
This repository contains an adapter, not a standalone 8B model. Load it on top of
Qwen/Qwen3-8B.
Project
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.
Adapter details
| 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.
Loading with Transformers and PEFT
1from transformers import AutoModelForCausalLM, AutoTokenizer
2from peft import PeftModel
3
4base_model_id = "Qwen/Qwen3-8B"
5adapter_id = "ThreePoints3/CPKeys-Qwen3-8B-RLOO-167"
6
7tokenizer = AutoTokenizer.from_pretrained(adapter_id)
8base_model = AutoModelForCausalLM.from_pretrained(
9 base_model_id,
10 torch_dtype="auto",
11 device_map="auto",
12)
13model = PeftModel.from_pretrained(base_model, adapter_id)
14model.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.
Evaluation summary
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:
- RLOO +167 performed strongly on several keyword and raw-prompt evaluations.
- RLOO +20 or +160 was stronger under some metrics and input modes.
- An SFT checkpoint was stronger on one small reference-reconstruction set.
See the
project report for tables, controls, and limitations.
Intended use
This adapter is intended for:
- research on competitive-programming problem-seed generation;
- generating candidate ideas for a human review queue;
- reproducing the CPKeys SFT/RLOO pipeline.
Generated outputs are not guaranteed to be correct, original, solvable, or contest-ready. They require similarity checking, technical validation, and human review.
Limitations
- Automatic quality and similarity scores are triage signals, not proof of problem quality or originality.
- Performance varies by input mode; this is not a universally best checkpoint.
- Intent normalization can route weak or non-generation requests incorrectly.
- The model may emit flawed constraints, underspecified tasks, solution leakage, or structures resembling existing problems.
- Native Qwen3 reasoning behavior and the CPKeys visible
<thinking> / <output> format must be handled consistently by the serving stack.
License
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.
