kdizzled commited on
Commit
23c2adf
·
verified ·
1 Parent(s): 3930e7d

Fix inconsistencies in README

Browse files
Files changed (1) hide show
  1. README.md +5 -11
README.md CHANGED
@@ -9,7 +9,7 @@ base_model:
9
 
10
  # **RocqStar Ranker Embedder**
11
 
12
- A self‑attentive embedding model for premise / proof selection in Rocq‑based interactive theorem proving (ITP). RocqStar is fine‑tuned from CodeBERT so that distances in the embedding space correlate with proof similarity rather than with surface‑level statement overlap. It replaces BM25/Jaccard similarity in retrieval‑augmented generation (RAG) pipelines such as **CoqPilot**, leading to higher proof success rates on the IMM‑300 benchmark.
13
 
14
  ## Model Details
15
 
@@ -19,7 +19,7 @@ RocqStar is a 125 M‑parameter Transformer encoder (768‑dim hidden size) wi
19
 
20
  * **Model type:** Transformer encoder with self‑attentive pooling
21
  * **Language(s):** Rocq / Coq (Gallina) syntax (tokens)
22
- * **License:**: Apache‑2.0
23
  * **Fine‑tuned from:** `microsoft/codebert-base`
24
 
25
  ### Model Sources
@@ -27,12 +27,6 @@ RocqStar is a 125 M‑parameter Transformer encoder (768‑dim hidden size) wi
27
  * **Repository:** [https://github.com/JetBrains-Research/big-rocq](https://github.com/JetBrains-Research/big-rocq)
28
  * **Paper:** (coming soon)
29
 
30
- ## Bias, Risks, and Limitations
31
-
32
- * **Domain limitation:** Trained only on Rocq projects; performance degrades on other languages or heavily tactic‑macro‑based proofs.
33
- * **Data leakage:** May memorise proofs present in training data and surface them verbatim.
34
- * **Proof similarity assumption:** While improved over BM25, cosine distance may still fail on very long or highly creative proofs.
35
-
36
  ## How to Get Started
37
 
38
  Go to [https://github.com/JetBrains-Research/big-rocq](https://github.com/JetBrains-Research/big-rocq). In subdirectory `ranker-server` you can find the server that you need to run to be able to use the model from the CoqPilot plugin. Also it may be used as a reference example of how to do inference with the model.
@@ -46,8 +40,8 @@ Go to [https://github.com/JetBrains-Research/big-rocq](https://github.com/JetBra
46
 
47
  ### Training Procedure
48
 
49
- * **Objective:** InfoNCE with temperature T = 0.07
50
- * **Batch size:** 512 pairs (effective) – gradient accumulation = 8×64
51
  * **Optimizer / LR:** AdamW, lr = 4e‑6, linear warm‑up 10 %, 22k steps
52
  * **Hardware:** 1× NVIDIA H100 GPU, 160 GB RAM, 14 h wall‑clock
53
 
@@ -71,4 +65,4 @@ Go to [https://github.com/JetBrains-Research/big-rocq](https://github.com/JetBra
71
 
72
  #### Summary
73
 
74
- RocqStar delivers consistent gains, up to 28% relative improvement over token‑overlap retrieval, especially for medium‑length theorems where proof similarity diverges most from statement similarity.
 
9
 
10
  # **RocqStar Ranker Embedder**
11
 
12
+ A self‑attentive embedding model for premise / proof selection in Rocq ITP. RocqStar is fine‑tuned from CodeBERT so that distances in the embedding space correlate with proof similarity rather than with surface‑level statement overlap. It replaces BM25/Jaccard similarity in retrieval‑augmented generation (RAG) pipelines such as **CoqPilot**, leading to higher proof success rates on the IMM‑300 benchmark.
13
 
14
  ## Model Details
15
 
 
19
 
20
  * **Model type:** Transformer encoder with self‑attentive pooling
21
  * **Language(s):** Rocq / Coq (Gallina) syntax (tokens)
22
+ * **License:** Apache‑2.0
23
  * **Fine‑tuned from:** `microsoft/codebert-base`
24
 
25
  ### Model Sources
 
27
  * **Repository:** [https://github.com/JetBrains-Research/big-rocq](https://github.com/JetBrains-Research/big-rocq)
28
  * **Paper:** (coming soon)
29
 
 
 
 
 
 
 
30
  ## How to Get Started
31
 
32
  Go to [https://github.com/JetBrains-Research/big-rocq](https://github.com/JetBrains-Research/big-rocq). In subdirectory `ranker-server` you can find the server that you need to run to be able to use the model from the CoqPilot plugin. Also it may be used as a reference example of how to do inference with the model.
 
40
 
41
  ### Training Procedure
42
 
43
+ * **Objective:** InfoNCE
44
+ * **Batch size:** 32
45
  * **Optimizer / LR:** AdamW, lr = 4e‑6, linear warm‑up 10 %, 22k steps
46
  * **Hardware:** 1× NVIDIA H100 GPU, 160 GB RAM, 14 h wall‑clock
47
 
 
65
 
66
  #### Summary
67
 
68
+ RocqStar delivers consistent gains, up to 28% relative improvement over Jaccard-index based retrieval, especially for medium‑length theorems where proof similarity diverges most from statement similarity.