vexp-ai/horizon-loras

Horizon LoRA adapters (math / code / science)

Three rank-64 LoRA adapters for DeepSeek-R1-Distill-Qwen-7B, trained on decontaminated, domain-filtered distilled traces. Part of the Horizon project: a verification-first layer for local LLMs, by the team behind vexp.

Read this first (honesty note)

We publish these adapters together with the measurement that they are mostly NOT the point. In our isolated-baseline protocol the fine-tuned specialists measured ~neutral on average: the system's value concentrates in hard verification, best-of-N selection and the agentic repair loop, not in the weights. Specifically:

• horizon-math-lora: +6 on MATH-500 in the v1.1 measurement, kept ON.
• horizon-code-lora: trained on competitive-programming style; it made HumanEval/MBPP slightly WORSE (-5/-1) and is disabled by default in the shipped config until retrained on function-completion style.
• horizon-science-lora: ~neutral.

We could have quietly shipped only the flattering one. The full numbers, protocol and per-run JSONs are in the repository; the honest read is the project's thesis: the model is not the product; the layer around it is.

Structure

One repo, three subfolders (horizon-math-lora, horizon-code-lora, horizon-science-lora), each a standard PEFT adapter (rank 64).

Usage

vLLM multi-LoRA serving (as used in the project):

vllm serve deepseek-ai/DeepSeek-R1-Distill-Qwen-7B \
  --enable-lora --max-lora-rank 64 \
  --lora-modules horizon-math-lora=<path>/horizon-math-lora \
                 horizon-code-lora=<path>/horizon-code-lora \
                 horizon-science-lora=<path>/horizon-science-lora

PEFT:

from peft import PeftModel
model = PeftModel.from_pretrained(base_model, "vexp-ai/horizon-loras", subfolder="horizon-math-lora")

Reproduce

Training pipeline (train/train_lora.py, data preparation and decontamination included) in the Horizon repository: about $12 of GPU time total on a rented 24 GB card.