thanhdath commited on
Commit
965a60b
·
verified ·
1 Parent(s): 9d5e7fa

add repo card

Browse files
Files changed (1) hide show
  1. README.md +51 -0
README.md ADDED
@@ -0,0 +1,51 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ---
2
+ license: other
3
+ tags:
4
+ - text-to-sql
5
+ - schema-linking
6
+ - grast-sql
7
+ language:
8
+ - en
9
+ ---
10
+
11
+ # GRAST-SQL bundle
12
+
13
+ Companion artifact bundle for the **GRAST-SQL** code repo
14
+ (`github.com/thanhdath/grast-sql-dev`). Holds the trained checkpoints, datasets,
15
+ precomputed embeddings, and a MongoDB dump that are too large for git, so another
16
+ machine can `git pull` the code and pull this bundle to continue work.
17
+
18
+ > See **`CONTINUE.md`** in the code repo for full bootstrap instructions.
19
+
20
+ ## Contents
21
+ ```
22
+ embeddings/
23
+ grast-0.6b-swift-merged/ # stage1 0.6B retriever (merged) — default
24
+ grast-4b-swift-merged/ # stage1 4B retriever (merged)
25
+ grast-0.6b-flagembed-merged/ # stage1 0.6B FlagEmbedding variant
26
+ gnn/
27
+ bird/*.pt spider/*.pt # stage2b GNN rerankers (0.6B/4B/8B, layer/hidden sweep)
28
+ deepseek-1.3b-l3-h2048/ # 1.3B GNN reranker (best epoch + train_log)
29
+ reranker/
30
+ deepseek-1.3b-lora/ # 1.3B reranker LoRA adapter (base: deepseek-coder-1.3b-base)
31
+ embeddings_cache/
32
+ bird-dev-reranker-1.3b/ # precomputed bird-dev embeddings (regenerable)
33
+ gretriever/ lora_v3_scoringhead-best.pt, lora_v5_birdheavy-best.pt
34
+ hetero_gnn/ v1t-epoch04.pt
35
+ data/ # mirrors the repo's data/ (BIRD dev/train graphs+emb, Spider 2.0)
36
+ mongo/grast/ # grast DB dump: synsql_samples / synsql_database / reforce (jsonl.gz)
37
+ ```
38
+
39
+ ## Quick start
40
+ ```bash
41
+ pip install -U "huggingface_hub[cli]"
42
+ hf download thanhdath/grast-sql-bundle --repo-type model --local-dir hf_bundle
43
+ # restore mongo:
44
+ python scripts/restore_grast_mongo.py --in hf_bundle/mongo/grast
45
+ ```
46
+
47
+ ## Notes
48
+ - Base models (Qwen3-Embedding/Reranker, deepseek-coder-1.3b) are **not** included —
49
+ download them from HF on demand; only fine-tuned/merged weights are shipped here.
50
+ - The `mats` MongoDB (eval result collections) and raw `synsql` (16 GB) are excluded;
51
+ they are regenerated by the eval scripts. The `grast` dump contains the training samples.