---
title: Feature Finder
emoji: 🔍
colorFrom: blue
colorTo: indigo
sdk: gradio
sdk_version: "5.50.0"
app_file: app.py
pinned: false
license: apache-2.0
hardware: h200
hf_oauth: false
---

# Feature Finder

Given a GitHub repo, surfaces a relevant recent arXiv paper for the team's next feature — paired with a PR-ready spec (summary, motivation, implementation plan, open questions) that a coding agent can pick up and run with.

For personalized recommendations trained on your team's full reading history, [sign up at engine.remyx.ai](https://engine.remyx.ai).

## About the model

The scoring model is a **LoRA adapter on Qwen3.5-2B**, distilled from a Pairwise Gaussian Process teacher trained on team-preference signals derived from GitHub merge histories.

- **Base**: [Qwen/Qwen3.5-2B](https://huggingface.co/Qwen/Qwen3.5-2B)
- **Adapter**: [`remyxai/mhpd-dpo-qwen3.5-2b-lora-v1.2-ipo-2epoch`](https://huggingface.co/remyxai/mhpd-dpo-qwen3.5-2b-lora-v1.2-ipo-2epoch) (current production; v1.5 in development)
- **Training objective**: IPO loss (β=0.1), bounded target gap that's more robust to noisy confident pairs than sigmoid DPO
- **Training data**: 50K preference pairs (`chosen` / `rejected`) generated from real GitHub repo merge histories
- **Eval**: candidate-disjoint repo-disjoint holdout — no paper appears in both train and eval — so accuracy reflects cross-repo generalization, not memorization

### Why a fine-tune vs zero-shot LLM?

A zero-shot LLM ranking arXiv papers against a team description gives generic "this paper looks topically relevant" answers. Our fine-tune learns the **preference signal** of which papers a team-aware reasoner (the GP teacher) actually flags as worth reading next. That extracts directional signal the LLM doesn't have access to.

The whole pipeline (GP teacher → preference pairs → LoRA distillation) is described in our training repository: [github.com/remyxai/mhpd-dpo-training](https://github.com/remyxai/mhpd-dpo-training).

### Personalized version

The open-source generalist you're using is trained on cross-repo preference signal. For per-team **specialist** models — adapted to your specific team's reading history via continued fine-tuning on your private preference data — sign up at [engine.remyx.ai](https://engine.remyx.ai). Specialists typically lift confidence by 10-15pts on team-specific picks vs the generalist.

## How it works

```
GitHub URL
  ↓
Fetch recent commits (GitHub Public API)
  ↓
Gemini Flash: filter noise + summarize commits into "experiment history" + "domain summary"
  ↓
sentence-transformers (all-MiniLM-L6-v2): embed domain summary, retrieve top-50 arXiv candidates
  ↓
v1.2 LoRA on Qwen3.5-2B: score logprob(paper | team_context) per candidate
  ↓
Confidence tier (z-score above pool median + embedding-cosine OOD floor)
  ↓
Base Qwen3.5-2B: draft a PR body integrating the paper
```

The arXiv candidate pool ([`remyxai/arxiv_pool_daily`](https://huggingface.co/datasets/remyxai/arxiv_pool_daily)) is precomputed daily — abstracts pre-embedded so per-request retrieval is sub-second CPU work. Categories covered: `cs.AI`, `cs.LG`, `cs.CL`, `cs.CV`, `cs.DB`, `cs.IR`, `cs.SE`, `stat.ML`.

## Calibration-aware confidence

The recommendation is tiered by **two** signals — z-score above the candidate-pool median AND embedding similarity to the team's domain summary. Both gates must clear for a tier:

| Tier | z-score | embedding cosine | Estimated accuracy |
|---|---|---|---|
| 🟢 **High** | ≥ 2.0σ | ≥ 0.60 | ~73% |
| 🟡 **Moderate** | ≥ 1.5σ | ≥ 0.50 | ~65% |
| 🟠 **Low** | ≥ 1.0σ | ≥ 0.40 | ~55% |
| 🔴 **Near-random** | otherwise | — | ~50% |

The two-signal design handles out-of-distribution repos honestly: a recommendation may be the clear winner *within* its candidate pool (high z-score) while the pool itself doesn't actually contain topical matches for the team's domain (low cosine). The OOD case correctly demotes to Low/Near-random and surfaces the upsell.

## Development

### Environment variables

| Variable | Purpose | Required |
|---|---|---|
| `HF_TOKEN` | Read access to the LoRA adapter + arxiv pool dataset | yes |
| `GEMINI_API_KEY` | Team-context summarization (Gemini Flash). Also used for PR-body generation when `MHPD_GENERATION_BACKEND=gemini` | yes |
| `MHPD_ADAPTER` | Override the LoRA adapter repo. Default: `remyxai/mhpd-dpo-qwen3.5-2b-lora-v1.2-ipo-2epoch` (v1.2). Will swap to v1.5 once trained. | no |
| `MHPD_GENERATION_BACKEND` | `local` (default) uses the base 2B on GPU for PR body draft. `gemini` routes PR body to Gemini Flash on CPU, freeing GPU time for the scoring step. | no |
| `USE_SPACES_GPU` | `true` (default) enables the `@spaces.GPU` decorator for Zero-GPU Spaces. Set `false` on persistent paid GPU Spaces. | no |
| `ARXIV_POOL_REPO` / `ARXIV_POOL_FILE` | HF dataset hosting the precomputed pool. Default: `remyxai/arxiv_pool_daily` | no |

### Local development

```bash
pip install -r requirements.txt
export HF_TOKEN=hf_...
export GEMINI_API_KEY=...
python app.py
```

### Architecture invariants

- **GPU only for scoring + PR body generation**. Embedding pre-filter, GitHub fetching, and Gemini call run on CPU before any GPU work is invoked.
- **arxiv_pool.parquet is precomputed** (daily cron via [GitHub Actions](https://github.com/remyxai/mhpd-dpo-training/blob/main/.github/workflows/refresh-arxiv-pool.yml)) — abstracts already embedded with all-MiniLM-L6-v2 to avoid embedding 500+ papers per cold start.
- **Single base model on GPU, LoRA toggled** via `enable_adapters()` / `disable_adapters()` for the scoring vs generation pass.

## License

Apache 2.0 for the Space code. The fine-tuned LoRA adapter is licensed per its model card on HF Hub. Underlying base model (Qwen3.5-2B) has its own license; arxiv abstracts are subject to arxiv terms.