---
license: other
tags:
  - text-to-sql
  - schema-linking
  - grast-sql
language:
  - en
---

# GRAST-SQL bundle

Companion artifact bundle for the **GRAST-SQL** code repo
(`github.com/thanhdath/grast-sql-dev`). Holds the trained checkpoints, datasets,
precomputed embeddings, and a MongoDB dump that are too large for git, so another
machine can `git pull` the code and pull this bundle to continue work.

> See **`CONTINUE.md`** in the code repo for full bootstrap instructions.

## Contents
```
embeddings/
  grast-0.6b-swift-merged/        # stage1 0.6B retriever (merged) — default
  grast-4b-swift-merged/          # stage1 4B retriever (merged)
  grast-0.6b-flagembed-merged/    # stage1 0.6B FlagEmbedding variant
gnn/
  bird/*.pt   spider/*.pt         # stage2b GNN rerankers (0.6B/4B/8B, layer/hidden sweep)
  deepseek-1.3b-l3-h2048/         # 1.3B GNN reranker (best epoch + train_log)
reranker/
  deepseek-1.3b-lora/             # 1.3B reranker LoRA adapter (base: deepseek-coder-1.3b-base)
embeddings_cache/
  bird-dev-reranker-1.3b/         # precomputed bird-dev embeddings (regenerable)
gretriever/  lora_v3_scoringhead-best.pt, lora_v5_birdheavy-best.pt
hetero_gnn/  v1t-epoch04.pt
data/                             # mirrors the repo's data/ (BIRD dev/train graphs+emb, Spider 2.0)
mongo/grast/                      # grast DB dump: synsql_samples / synsql_database / reforce (jsonl.gz)
```

## Quick start
```bash
pip install -U "huggingface_hub[cli]"
hf download thanhdath/grast-sql-bundle --repo-type model --local-dir hf_bundle
# restore mongo:
python scripts/restore_grast_mongo.py --in hf_bundle/mongo/grast
```

## Notes
- Base models (Qwen3-Embedding/Reranker, deepseek-coder-1.3b) are **not** included —
  download them from HF on demand; only fine-tuned/merged weights are shipped here.
- The `mats` MongoDB (eval result collections) and raw `synsql` (16 GB) are excluded;
  they are regenerated by the eval scripts. The `grast` dump contains the training samples.