--- license: apache-2.0 language: en base_model: Qwen/Qwen3-8B library_name: mlx pipeline_tag: text-generation tags: [sports, experimental, qwen3, mlx, lora] --- # Relay Sports 8B — trained experimental preview 0.1 A local sports Q&A experiment by Yoseph Bagighni. This package contains genuinely fine-tuned weights adapted from **Qwen3-8B**, plus a local Python runner. It is not a foundation model trained from scratch. The smaller-model experiments were unsuccessful; this first trained desktop preview uses an 8B base. **This is a research preview with known errors.** It has no demonstrated overall accuracy improvement over its base, and it is not ready to serve as a dependable coach or health product. ## Run it on an Apple-silicon Mac Tested on an M1 Pro Mac with 16 GiB memory, macOS 27.2 and Python 3.11.17. Windows, Linux and physical iPhones were not tested. The download is about 5.1 GB and already includes the model weights; no Ollama base-model pull or API key is required. 1. Extract the ZIP and open Terminal in its `Relay-Sports-8B` folder. 2. Use Python 3.11 or newer to create the environment and install the pinned libraries: ```sh python3.11 -m venv .venv .venv/bin/python -m pip install -r requirements.txt ``` 3. Ask one question: ```sh .venv/bin/python ask.py "What is the total distance of 6 x 75 meters?" ``` Run `.venv/bin/python ask.py` for interactive questions; enter `/bye` to exit. Ctrl-C stops the process. Each question is independent. To supply your own short UTF-8 note, use `.venv/bin/python ask.py --notes your-notes.txt "What time was recorded?"`. The CLI reads only the explicitly selected note, capped at 8192 bytes. Questions are capped at 4096 UTF-8 bytes, and the actual templated prompt plus 384 output tokens must fit 4096 tokens. Longer material is rejected. Output can be cut off at the generation limit. ## What was actually trained Starting conversion: `mlx-community/Qwen3-8B-4bit`, revision `545dc4251c05440727734bcd94334791f6ab0192`. Its original weight checksum matched the pinned public conversion before training. - Upstream foundation: [Qwen/Qwen3-8B](https://huggingface.co/Qwen/Qwen3-8B). - Data: 92 original/code-generated training examples and 11 same-skill validation examples. They cover basic arithmetic, elementary sports definitions, fictional logs, identity and professional boundaries. No personal notes, real health records, evaluation fixtures or teacher-model answers were training inputs. - Method: 92 steps, batch 1, two final layers, LoRA rank 4 / scale 4, learning rate 5e-6, masked prompts, maximum 256 training tokens, seed 20261114. Approximately 0.500 million parameters were trainable within an 8.19 billion-parameter model. - Export: most original layers remain 4-bit; trained projections are fused at higher precision. The complete export was reloaded and tested. Sampled comparison against the dequantized base confirmed 33,340 changed values among 65,536 inspected projection values; all ten learned LoRA B matrices were nonzero and finite. Serialization alone is not the change proof. ## Checks and limitations Both the unchanged base and the exported checkpoint received the same system prompt, notes formatting, greedy decoding, seed 42, thinking disabled and 384-token output limit. - Development screen: base 24/28, trained 25/28 automated passes. These cases were used during development. The extra pass reflects citation-bracket formatting, not demonstrated factual improvement. All six development calculations were correct. - Previously unrun 26-question synthetic screen: **23/26 automated passes for both models**. This is not an accuracy estimate for general sports or health questions. - A correct abstention on conflicting notes was rejected by a strict pattern check. Full-answer review also found errors that the pattern checks missed. - Both models incorrectly calculated 1750 yards as 1604.4 meters; the correct result is 1600.2 meters. Seven of eight final calculation questions were correct. - Both followed a hostile footer inside a supplied note and returned the wrong venue. Treat this as a known prompt-injection limitation. - Additional explanations confused SCM/LCM terminology or added unsupported details about starting position. Do not assume every statement in an otherwise correct answer is accurate. - An extra CLI check misread a compact `time1:18.40` field as `18.40`. Clearly spaced fields passed the repeated functional check, but this does not establish reliable reading of arbitrary notes. - The two clinical test prompts received refusals/referrals, but that small sample is not clinical or safety validation. Do not use this model for diagnosis, medication decisions or medical clearance. Codex and Gemini inspected the full answers. There was no clinician/coach validation or broad sports benchmark. No further training used the final 26 answers; a future version needs new unseen tests. `CHECKS.json` contains the exact release identity, results, review notes and limitations. `SHA256SUMS.txt` covers the packaged files. Peak MLX allocation measured 5.151 GiB in the final synthetic run; total system memory use is higher. This is a Mac measurement, not an iPhone requirement or a minimum-memory guarantee. ## Privacy and license After installation/download, the supplied runner loads the adjacent model and sets Hugging Face/Transformers offline flags. It has no cloud fallback, transcript saving, telemetry code, vault indexing, tools or autonomous actions. Installing dependencies and obtaining the package require network access; entered questions and explicitly selected notes are processed locally by the runner. The upstream Apache License 2.0 and attribution are included in `LICENSE` and `NOTICE`. Preserve them when redistributing. The Relay name, code and adaptation do not imply exclusive ownership of the upstream foundation weights. An affordable fitness app and a reliable phone-size model remain future work.