guychuk commited on
Commit
3a72ccf
·
verified ·
1 Parent(s): 58cc33b

Upload README.md with huggingface_hub

Browse files
Files changed (1) hide show
  1. README.md +49 -0
README.md ADDED
@@ -0,0 +1,49 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ---
2
+ license: mit
3
+ language:
4
+ - en
5
+ base_model: Qwen/Qwen2.5-Coder-1.5B-Instruct
6
+ tags:
7
+ - code
8
+ - verus
9
+ - formal-verification
10
+ - fault-localization
11
+ - energy-based-model
12
+ - lora
13
+ - peft
14
+ library_name: peft
15
+ ---
16
+
17
+ # VericodingEBM — Per-Line Fault Localizer for Verus
18
+
19
+ LoRA + per-line scalar head trained on top of [Qwen2.5-Coder-1.5B-Instruct](https://huggingface.co/Qwen/Qwen2.5-Coder-1.5B-Instruct) to score every line of a Verus implementation with an energy proxy for *"this line is the bug."*
20
+
21
+ Submitted to the **Apart × Atlas Computing Secure Program Synthesis Hackathon, Track 3 (Vericoding)**.
22
+
23
+ 📄 **Paper:** see [`paper/main.pdf`](https://github.com/ozlabsai/VericodingEBM/blob/main/paper/main.pdf)
24
+ 💾 **Code + reproducibility:** https://github.com/ozlabsai/VericodingEBM
25
+ 📊 **Training data:** [`OzLabs/VericodingEBM-data`](https://huggingface.co/datasets/OzLabs/VericodingEBM-data)
26
+
27
+ ## Headline results (run #10, this checkpoint)
28
+
29
+ | Measurement | This model | Best frontier LLM |
30
+ |---|---|---|
31
+ | Per-line top-3 recall on Verus dev-test (n=609 FAILs) | **0.84** | 0.74 (Claude Opus 4.7) |
32
+ | Whole-impl discrimination AUROC | 0.78 | **0.91** (GPT-5.5) |
33
+ | Closed-loop CEGIS repair@1 (n=100) | 25% | **30%** (LLM self-judged) |
34
+
35
+ ## What's in this repo
36
+
37
+ - `adapter/` — LoRA adapter (PEFT format, rank 16, alpha 32, embed_lora_rank 8) for Qwen2.5-Coder-1.5B-Instruct
38
+ - `head.pt` — per-line scoring head weights (small MLP over sentinel-token hidden states)
39
+ - `scalar_head.pt` — whole-impl attention-pool head weights
40
+
41
+ To run inference you need all three files plus the training code at https://github.com/ozlabsai/VericodingEBM.
42
+
43
+ ## Marker-leak audit (paper §4.6)
44
+
45
+ This checkpoint is **marker-AVERSE**: its per-line top-1 recall jumps from 4% to 56% when the `// FAILS` debug markers are stripped from the input (delta = −52pp). This is the result of the counterfactual-marker augmentation described in paper §B; pre-fix checkpoints (run #7) are marker-RELIANT (signal collapses without markers).
46
+
47
+ ## License
48
+
49
+ MIT (see [GitHub repo](https://github.com/ozlabsai/VericodingEBM)).