Llama-3.2-3B-HYROX
A LoRA adapter for sports performance commentary
2 min readupdated
This compact LoRA (a small add-on over a base model) transforms the Llama-3.2-3B language model into a specialized competition analyst for HYROX events. Based on the provided table of race results, split times, and age group standings, the model generates a short, coherent text in the second person. It highlights the athlete's strengths and weaknesses, explains where time was gained or lost, and relies strictly on facts from the given prompt. The solution was trained using Supervised Fine-Tuning on 1,644 high-quality texts and demonstrates an extremely low error rate.
Key features
- High accuracy: the hallucination rate is reduced to 1.4% compared to 24.1% for the untrained base model.
- Architecture efficiency: uses LoRA with rank 16 across all linear layers, affecting only 0.75% of the weights (24.3M parameters, file size 97 MB).
- Strict format: the model works exclusively with result tables of a specific structure and maintains a text length of up to 110 words.
What you can do
- Report generation: automate the creation of analytical summaries and personal commentary following HYROX competitions.
- Sports analysis: quickly process tabular split times and extract key patterns from an athlete's performance.
Example
import torch
from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer
BASE = "meta-llama/Llama-3.2-3B-Instruct"
ADAPTER = "vladmateei/Llama-3.2-3B-HYROX-Commentary-LoRA"
tokenizer = AutoTokenizer.from_pretrained(BASE)
base = AutoModelForCausalLM.from_pretrained(BASE, dtype=torch.bfloat16).to("cuda")
model = PeftModel.from_pretrained(base, ADAPTER).eval()Read more
Details• Benchmarks: the hallucination rate on held-out tests was 1.4%, and the Insight analytical value metric was 0.82.
• System requirements: requires access to the base Llama-3.2-3B-Instruct model and about 7 GB of VRAM (bfloat16).
• Stack: peft, transformers, torch.
• Limitations: the model is not a universal chatbot, does not give training advice, and works strictly with a fixed input table format.










Comments