Paper | Website | Leaderboard | Dataset | Google Colab

FreshStack is a modular framework to automatically build realistic IR/RAG benchmarks from niche, community-sourced technical content (e.g., Stack Overflow + GitHub repositories). It supports:

• Scraping human-asked queries based on StackOverflow.
• Gathering up-to-date corpora via chunking any GitHub repository.
• Retrieval evaluation of any dense/multi-vector model on the Freshstack repository.
• Datasets released under CC-BY-SA 4.0 and code and scripts under Apache 2.0 License.

Installation

Install via pip, tested with Python 3.10+:

pip install freshstack

If you want to build from source, use:

git clone https://github.com/fresh-stack/freshstack.git
cd freshstack
pip install -e .

🚀 Quickstart: Load Freshstack Dataset

from freshstack.datasets import DataLoader

dataloader = DataLoader(
    queries_repo="freshstack/queries-oct-2024", 
    corpus_repo="freshstack/corpus-oct-2024",
    topic="langchain") # or "yolo", "angular", "laravel" or "godot"

# Loads the corpus, queries and nuggets in the BEIR format
corpus, queries, nuggets = dataloader.load(split="test")

# Loads the qrels (nuggets), qrels (query) and query to nugget mapping
qrels_nuggets, qrels_query, query_to_nuggets = dataloader.load_qrels(split="test")

🚀 Quickstart: Model Evaluation

1. Evaluate only the retrieved results

# Your runfile can be stored as a .txt in the following format: [qid, Q0, docid, 0, score, run_name], e.g.,
# 76185522 Q0 angular/adev/src/content/tutorials/learn-angular/steps/14-routerLink/answer/src/app/app.component.ts_0_368 0 0.7353782057762146 your_model_name

from freshstack import util
from freshstack.retrieval.evaluation import EvaluateRetrieval

# retrieval_results: dict[str, dict[str, str]] with qid: {doc_id: score}
retrieval_results = util.load_runfile("<path_to_your_runfile>")
evaluator = EvaluateRetrieval(k_values=[10, 20, 50])
alpha_ndcg, coverage, recall = evaluator.evaluate(
    qrels_nuggets=qrels_nuggets,
    query_to_nuggets=query_to_nuggets,
    qrels_query=qrels_query,
    results=retrieval_results,
)

2. Evaluate any dense embedding model (e.g., Qwen3-0.6B-embedding) using BEIR.

Make sure you install the latest PyPI BEIR repository: pip install beir

from beir.retrieval import models
from beir.retrieval.evaluation import EvaluateRetrieval as BEIREval
from beir.retrieval.search.dense import DenseRetrievalExactSearch as DRES
from freshstack.retrieval.evaluation import EvaluateRetrieval

# Custom query prompt for evaluating the Qwen3-0.6B model on Freshstack.
query_prompt = "Instruct: Given a technical question, retrieve relevant code snippets or technical documentation that best answer the question\nQuery: "

model = DRES(models.SentenceBERT(
    "Qwen/Qwen3-Embedding-0.6B",
    max_length=2048, # IMP: keep the max_length for both query & passage atleast 2048 tokens.
    prompts={"query": query_prompt, "passage": ""},
    model_kwargs={
        "attn_implementation": "flash_attention_2", 
        "device_map": "auto", 
        "torch_dtype": "bfloat16"
    },
    tokenizer_kwargs={"padding_side": "left"},
), batch_size=32)

retriever = BEIREval(model, score_function="cos_sim")
retrieval_results = retriever.retrieve(corpus=corpus, queries=queries)

# Evaluate and compute retrieval score once you have results
evaluator = EvaluateRetrieval(k_values=[10, 20, 50])
alpha_ndcg, coverage, recall = evaluator.evaluate(
    qrels_nuggets=qrels_nuggets,
    query_to_nuggets=query_to_nuggets,
    qrels_query=qrels_query,
    results=retrieval_results,
)

3. Evaluate any multi-vector model (e.g., ColBERT) using Pylate.

Make sure you install the latest PyLate repository: pip install pylate.

from pylate import indexes, models, retrieve
from freshstack.retrieval.evaluation import EvaluateRetrieval

# Step 1: Load the ColBERT model
model = models.ColBERT(
    model_name_or_path="lightonai/GTE-ModernColBERT-v1",
    query_length=2048, document_length=2048
)

# Step 2: Initialize the Voyager index or (PLAID index)
index = indexes.Voyager(
    index_folder=f"./langchain_index",
    index_name="index",
    override=False,  # This overwrites the existing index if any
)

# Step 3: Encode the documents and add them to index
documents_ids = list(corpus.keys())
documents_embeddings = model.encode(
    [doc["text"] for doc in corpus.values()],
    batch_size=32,
    is_query=False,  # Ensure that it is set to False to indicate that these are documents, not queries
)

index.add_documents(
    documents_ids=documents_ids,
    documents_embeddings=documents_embeddings,
)

# Step 5: Compute query embeddings
query_ids = list(queries.keys())
queries_embeddings = model.encode(
    list(queries.values()),
    batch_size=32,
    is_query=True,  # Ensure that it is set to False to indicate that these are queries
)

# Step 6: Initialize the ColBERT retriever with the Voyager index & retrieve documents
retriever = retrieve.ColBERT(index=index)
scores = retriever.retrieve(
    queries_embeddings=queries_embeddings,
    k=50,  # Retrieve top-k results based on the maximum k value specified
    batch_size=1,  # We have kept a batch size of 1 to avoid memory issues.
    device="cpu",  # Use CPU for inference, change to "cuda" if you have a GPU available.
)

# Step 7: Prepare the results in the required BEIR format
retrieval_results = {}
for query_id, doc_scores in zip(query_ids, scores):
    retrieval_results[query_id] = {}
    for doc_id, score in doc_scores:
        retrieval_results[query_id][doc_id] = score

# Step 8: Evaluate and compute retrieval score once you have results
evaluator = EvaluateRetrieval(k_values=[10, 20, 50])
alpha_ndcg, coverage, recall = evaluator.evaluate(
    qrels_nuggets=qrels_nuggets,
    query_to_nuggets=query_to_nuggets,
    qrels_query=qrels_query,
    results=retrieval_results,
)

📚 Raw Freshstack Datasets (Oct 2024)

The raw freshstack datasets can be downloaded via HF:

• langchain, yolo, laravel, angular, godot (via huggingface.co)

from datasets import load_dataset
queries = load_dataset("freshstack/queries-oct-2024", subset="yolo")
corpus = load_dataset("freshstack/corpus-oct-2024", subset="yolo")

🧭 Project Structure

freshstack/
├─ examples/            # contains examples
│   ├─ chunking/        # examples for github repo chunking
│   ├─ evaluation/      # examples for model eval on freshstack
├─ freshstack/          # core logic modules
│   ├─ retrieval/       # code for retrieval evaluation
│   ├─ datasets/        # code for freshstakc dataloader
│   └─ chunking/        # code for github repo chunking
└─ pyproject.toml

FreshStack Leaderboard

The upto date leaderboard for Freshstack (version oct-2024) is provided here: https://fresh-stack.github.io/#leaderboard.

NOTE: Below is the snapshot of the Freshstack leaderboard from Jun 12th 2025.

👥 Contribute your model to the leaderboard

1. Fork the repo (https://github.com/fresh-stack/fresh-stack.github.io)
2. Create a new branch (git checkout -b <your_branch>)
3. Add your model scores in the following format to the leaderboard_data.json:

{
    "leaderboardData": [
        {
            "info": {
                "name": "BM25",
                "size": "-",
                "type": "open_source",
                "date": "2024-11-01",
                "link": "https://github.com/castorini/pyserini"
            },
            "datasets": {
                "langchain": {"alpha_ndcg_10": 0.230, "coverage_20": 0.475, "recall_50": 0.261},
                "yolo":      {"alpha_ndcg_10": 0.137, "coverage_20": 0.342, "recall_50": 0.337},
                "laravel":   {"alpha_ndcg_10": 0.319, "coverage_20": 0.602, "recall_50": 0.441},
                "angular":   {"alpha_ndcg_10": 0.259, "coverage_20": 0.551, "recall_50": 0.340},
                "godot":     {"alpha_ndcg_10": 0.144, "coverage_20": 0.268, "recall_50": 0.200},
                "average":   {"alpha_ndcg_10": 0.218, "coverage_20": 0.448, "recall_50": 0.316},
            }
        },
        ...
    ]
}

4. Submit a pull request, ideally including:

   • The updated leaderboard_data.json
   • Pipeline invocation script (reference)
   • Brief evaluation summary (reference)

All contributions welcome—especially new domain expansions, evaluation improvements, and retrieval baselines!

📄 Citation

If you use FreshStack in your work, please cite:

@article{thakur-freshstack:2025,
  author       = {Nandan Thakur and
                  Jimmy Lin and
                  Sam Havens and
                  Michael Carbin and
                  Omar Khattab and
                  Andrew Drozdov},
  title        = {FreshStack: Building Realistic Benchmarks for Evaluating Retrieval
                  on Technical Documents},
  journal      = {CoRR},
  volume       = {abs/2504.13128},
  year         = {2025},
  url          = {https://doi.org/10.48550/arXiv.2504.13128},
  doi          = {10.48550/ARXIV.2504.13128},
  eprinttype    = {arXiv},
  eprint       = {2504.13128},
  timestamp    = {Thu, 22 May 2025 21:00:35 +0200},
  biburl       = {https://dblp.org/rec/journals/corr/abs-2504-13128.bib},
  bibsource    = {dblp computer science bibliography, https://dblp.org}
}

The main contributors of this repository are:

• Nandan Thakur, Personal Website: thakur-nandan.gitub.io

Contact person: Nandan Thakur, nandant@gmail.com

Don't hesitate to send us an e-mail or report an issue, if something is broken (and it shouldn't be) or if you have further questions.

This repository contains experimental software and is published for the sole purpose of giving additional background details on the respective publication.

Collaboration

This project is developed in collaboration with the following organizations:

Organization	Logo
University of Waterloo
Databricks