rotowire-gemma-4-lora
LoRA adapters for turning NBA reports into tables
2 min readupdated
A set of five LoRA adapters (small adjustments on top of a base model) for the Gemma 4 family (12B, E4B, and E2B models) that turn an NBA game text report into two structured tables with team and player statistics. The adapters were trained on the Rotowire dataset using improved annotations from Opus and show high table generation accuracy (up to 97.1 F1 for players). The tool will be useful for developers and researchers who need to extract structured data from sports or text reports using local language models.
Key features
- Rotowire specialization: adapts text reviews of NBA matches into Team and Player table formats.
- High accuracy: top versions reach 95.7 (Team) and 97.1 (Player) on the cell F1 metric.
- Three base sizes: options for Gemma 4 models with 12B, E4B, and E2B sizes.
- Ready-made settings: includes options based on different training annotations (Opus relabels and Wu et al.).
What you can do
- Run via vLLM: deploy the model locally with support for LoRA modules to process requests.
- Generate box scores: automatically convert text reports of sports matches into ready-made Markdown tables.
- Compare approaches: test the difference between adapter versions trained on different data annotations.
Example
from pathlib import Path
from openai import OpenAI
client = OpenAI(base_url="http://127.0.0.1:8000/v1", api_key="-")
reply = client.chat.completions.create(
model="e4b-teacher-r32",
temperature=0,
max_completion_tokens=1536,
messages=[
{"role": "system", "content": Path("schema-v1-prompt.txt").read_text()},
{"role": "user", "content": "Report:\n" + report_text},
],
)
print(reply.choices[0].message.content)Read more
DetailsBenchmarks: up to 95.7 (Team) and 97.1 (Player) cell F1 on the Rotowire test set.
Training: LoRA rank 32, alpha 32, dropout 0.05 on attention and MLP projections, 2 epochs.
Limitations: tested only on a single domain (NBA reports) and a fixed set of 47 columns. Requires strict adherence to the prompt template and decoding via vLLM.










Comments