jsantillana/qwen2.5-3b-f1-strategy-lora

Qwen2.5-3B F1 Strategy LoRA (grounded, coverage-aware)

LoRA adapter for Qwen/Qwen2.5-3B-Instruct, fine-tuned to generate Formula 1 strategy explanations grounded in telemetry-derived structured data. Companion model to the paper "Precision Is Not Faithfulness: Coverage-Aware Evaluation of Grounded Generation with a Complete Oracle."

In the paper this fine-tuned 3B model attains the highest precision (faithfulness) of all systems evaluated while staying substantive (more claims per instance than the terse frontier models) — i.e. its high faithfulness is not bought by abstention.

Usage

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-strategy-lora")

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/o_proj), 3 epochs, lr 2e-4.
• Data: data/structured/sft.jsonl in the project repo (grounded explanations).

Links

• Code: https://github.com/vectrayx/precision-is-not-faithfulness
• Dataset: https://huggingface.co/datasets/jsantillana/f1-strategy-faithfulness
• Demo: https://huggingface.co/spaces/jsantillana/faithful-strategy-engineer-f1

Honest caveats

• Silver supervision: training targets are deterministic faithful templates, which favors a grounded, concise style and risks rewarding template mimicry. See the paper's Limitations.
• This adapter is recipe-identical to the paper's run but, due to training nondeterminism, not bit-identical to the exact checkpoint behind the reported numbers.
• Raw F1/FOM data is not redistributed; only derived structured data and code are released.