qwen-2.5-1b-rlcd
Chinese local MLX demo that scores 2-5 candidate decisions with a small Qwen model and drives a Snake game by choosing among moves.
Overview
This repository is a GitHub derivative of harshatheg/Qwen-2.5-1B-RLCD on Hugging Face, adapted for Apple Silicon and MLX with Chinese documentation. The actual model is mlx-community/Qwen2.5-1.5B-Instruct-4bit (weights not included). Changes include offline start and download scripts, a decision_demo.py that takes a custom question and 2-5 single-token candidate labels and returns candidate probabilities with conflict detection, a smoke test, and pinned dependencies.
A web page shows structured decisions, and a Snake Lab page lets the local model choose between straight, left and right at each step, with planning assistance that filters infeasible moves first, collision protection, and live display of probabilities and latency. The README is explicit that the probabilities are normalized model scores within the candidate set, not calibrated correctness, and that the upstream web page applies a heuristic 0.75 floor in some multi-token conflicts.
What the decision model decides
| Type | Question | Answers |
|---|---|---|
| Choice | User says the software does not work. What next action should be taken? | A · B · C |
| Choice | Which move should the snake make: straight, left or right? | straight · left · right |
How it works
- Download the 4-bit MLX model once
- Build a prompt with the question and single-token labeled options
- Read normalized probabilities for each candidate from one pass
- In Snake Lab, filter unsafe moves with planning, then let the model choose
Strengths
- Runs fully offline on Apple Silicon after download
- Shows per-candidate probabilities and latency live
- Conflict checks avoid the upstream heuristic branch
Caveats
- Probabilities are not calibrated against real data, as the README states
- Small model can loop in Snake without planning help
- Mac-only (MLX)
Getting started
python3.12 -m venv .venv && .venv/bin/python -m pip install -r requirements-lock.txt && ./download-model.sh && ./start.shAt a glance
- Type
- Model
- Status
- prototype
- Language
- Python
- Runtime
- MLX on Apple Silicon, local web page
- Has code
- yes
- Has tests
- yes
- License
- Apache-2.0
Use cases and integrations
How curated lists describe it
Local MLX decision probability demo derived from harshatheg/Qwen-2.5-1B-RLCD on…
wh000wh000/awesome-jev-live
Frequently asked questions
What does it decide?
Which of 2-5 labeled candidate actions fits a question, and which way a Snake game should turn.
Are the probabilities calibrated?
No; the README says they are normalized scores within the candidate set.
Does it need a TypeSafe key?
No; it runs a local Qwen model through MLX.
Related projects
Research engine for local System One decisions via llama.cpp logits projection with Qwen, running sub-second on a 4 GB laptop GPU.
ModelresearchPythonJev-inspired technique scoring single-token labels with an LLM's head via vLLM to fill typed fields, with a fast Mario demo.
BenchmarkresearchPythonllama.cpp-based C++ server that turns any LLM into a System One decision model, reading answers from next-token logits.
ProjectbetaC++Python library that imitates Jev-style typed decisions (choice, score, yes/no) on local LLMs via Ollama, vLLM or llama.cpp.
ProjectprototypePythonMLX library that turns an existing local model into a decision engine by reading candidate logits, with no fine-tuning, on Apple Silicon.
ProjectprototypePythonTurns any local LLM into a Jev-style engine via masked-logit softmax, or proxies OpenAI-compatible APIs or Jev; includes a chat crush radar.
ServiceprototypePython