CoolFace
Modelpublic

jsantillana/qwen2.5-3b-f1-commentary-lora-v2

sourceHugging Faceotherupdated 4mo agoView on Hugging Face
0likes8downloads
Model Card

Qwen2.5-3B F1 Live Commentary LoRA (EN/ES/PT)

LoRA adapter for Qwen/Qwen2.5-3B-Instruct, fine-tuned to generate short (2-3 sentence), grounded live F1 race commentary from a structured live-state snapshot (positions, gaps, tyre compounds/age, on-track battles, pace deltas, safety-car flag). Powers the live commentary feature of Pitwall.

Companion adapter to jsantillana/qwen2.5-3b-f1-strategy-lora (the paper's post-race strategy-explanation model, fine-tuned on a different prompt/format). This adapter is fine-tuned specifically on the prompt/context format used by services/api/commentary.py::build_context and SYSTEM_PROMPT/PROMPTS, so it learns to terminate cleanly after 2-3 sentences instead of degenerating into repetition.

Usage

python
from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer
base = "Qwen/Qwen2.5-3B-Instruct"
tok = AutoTokenizer.from_pretrained(base)
model = AutoModelForCausalLM.from_pretrained(base, torch_dtype="auto", device_map="auto")
model = PeftModel.from_pretrained(model, "jsantillana/qwen2.5-3b-f1-commentary-lora-v2")

Training

  • —Base: Qwen/Qwen2.5-3B-Instruct (Qwen Research License — this adapter inherits it; non-commercial).
  • —Method: LoRA (r=16, alpha=32, dropout=0.05; q/k/v/oproj), 3 epochs, lr 2e-4, batch 2 x gradaccum 8, bf16.
  • —Data (v2): 2504 examples merged from (a) a trilingual (EN/ES/PT) GPT-4o-synthesized set, prompt-matched to commentary.py's context/prompt and grounded in the live track-state store (strategy_sim/track_store.state_at_lap), and (b) real F1TV broadcast commentary (EN/ES subtitle transcripts) filtered to examples whose claims verify against the same state store (src/eval/live_faithfulness.py). 2025/2026 seasons held out for eval.

Links

  • —Code: https://github.com/vectrayx/precision-is-not-faithfulness
  • —Product: https://pitwall.jsantillana.com

Honest caveats

  • —Silver supervision: training targets are deterministic grounded templates, so the model may lean toward template-like phrasing rather than freer commentary style.
  • —Raw F1/FOM data is not redistributed; only derived structured data and code are released.