Top AI Repos — open-source AI, indexed and scored
Top AI Repos tracks AI repositories on GitHub and answers two different questions about each one: is it moving right now, and would you bet a product on it.
Top AI Repos tracks AI repositories on GitHub and answers two different questions about each one: is it moving right now, and would you bet a product on it.
🚀 Ultra Recipe for Training Long-Horizon Search Agents - matching frontier AI's search capability with a 20B model + stateful harness
| Date | Stars |
|---|---|
| 2026-07-31 | 897 |
| 2026-08-03 | 922 |
| 2026-08-06 | 922 |
Today
— stars today
This week
— stars this week
This month
— stars this month
Momentum
0.0
growth rate 0.00%/day
# Harness-1
[](https://github.com/pat-jj/harness-1/blob/main/inference/tinker_inference.md)
[](https://huggingface.co/pat-jj/harness-1)
[](https://arxiv.org/abs/2606.02373)
[](https://x.com/patpcj/status/2063298457398636570?s=20)
Harness-1 is a 20B search agent trained with reinforcement learning inside a
stateful retrieval harness. The harness maintains recoverable search state:
candidate documents, curated evidence, evidence links, verification records, and
budget-aware context. The policy keeps the semantic decisions: what to search,
which documents to inspect or curate, what claims to verify, and when the
evidence is sufficient.
## Quickstart
For a minimal local smoke test, you need:
- Linux with Python `3.11+`.
- `uv` installed.
- A CUDA-compatible NVIDIA GPU environment.
- vLLM with GPT-OSS support.
- Access to the released Hugging Face checkpoint.
Install dependencies:
```bash
uv sync --extra vllm
```
Set the checkpoint:
```bash
export HARNESS1_HF_MODEL=pat-jj/harness-1
```
Start with the detailed vLLM and BrowseComp+ guide:
```bash
less docs/run_vllm_browsecompplus.md
```
## Model Checkpoint
The released Harness-1 weights are hosted on Hugging Face:
```text
https://huggingface.co/pat-jj/harness-1
```
vLLM downloads the weights from Hugging Face on first use and then reuses the
local Hugging Face cache. See the Hugging Face model page for model-card details,
usage restrictions, and checkpoint metadata.
## Training Data And Corpora
The training data used for Harness-1 is published at
[`pat-jj/harness-1-train-data`](https://huggingface.co/datasets/pat-jj/harness-1-train-data).
It contains one `train` split with a `stage` column:
- `sft`: 899 raw GPT-5.4-generated v8d SFT trajectories from
`generate_sft_ultra_0417.py`, used by `train_sft_ultra_0417.py`.
- `rl`: 3,453 SEC training-split query records used for RL
(`TRAIN_DATASETS=sec`, `RL_QUERY_SPLIT=train`).
```python
from datasets import load_dataset
ds = load_dataset("pat-jj/harness-1-train-data", split="train")
sft = ds.filter(lambda row: row["stage"] == "sft")
rl = ds.filter(lambda row: row["stage"] == "rl")
```
The same dataset repo also includes the retrieval corpora under `corpora/`, with
chunk text and cleaned metadata for BrowseComp+, web, patents, and SEC. For
example:
```python
from datasets import load_dataset
sec_corpus = load_dataset(
"parquet",
data_files="hf://datasets/pat-jj/harness-1-train-data/corpora/sec/train/*.parquet",
split="train",
)
```
## What You Can Do
- Serve the released checkpoint locally with vLLM.
- Run raw `/v1/completions` smoke tests with token-id outputs.
- Evaluate Harness-1 search behavior on BrowseComp+ when a compatible retrieval
backend is available.
- Run Tinker-hosted inference with the published checkpoint.
- Inspect and extend the stateful search harness, tool environment, training
scripts, and evaluation runners.
- Run ablations and baselines for supported datasets.
## Repository Layout
- `docs/`: user-facing guides and runbooks.
- `harness/`: shared search harness, tools, trajectory, task, reranking, and
configuration modules.
- `inference/`: Harness-1 evaluation, component ablations, HF inference, and
vLLM inference utilities.
- `inference/baselines/`: in-domain and transfer baseline evaluation runners.
- `training/`: SFT data generation, SFT training, RL training, and launch scripts.
- `datagen/` and `eval_scripts/`: dataset and auxiliary evaluation code.
- `model_export/`: helper scripts for merging a private Tinker adapter into aExcerpt of 10,522 characters
Read on GitHub
Patrick Jiang · UIUC
23
1
Would you bet a product on this? Bounded 0–100 and slow moving.
matched fp:d4c5eadf814c8416, llm:Description: 'Ultra Recipe for Training Long-Horizon Search Agents - matching frontier AI's search capability with a 20B model + stateful harness' (Python).
matched fp:d4c5eadf814c8416, llm:Description: 'Ultra Recipe for Training Long-Horizon Search Agents - matching frontier AI's search capability with a 20B model + stateful harness' (Python).
matched fp:d4c5eadf814c8416, llm:Description: 'Ultra Recipe for Training Long-Horizon Search Agents - matching frontier AI's search capability with a 20B model + stateful harness' (Python).