Hugging Face's logo

Zishan-Shao
/
decodeshare

interpretability
mechanistic-interpretability
activation-steering
llm
kv-cache
icml-2026
Model card Files
xet
 Community
decodeshare
File size: 1,785 Bytes
0a40186
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
87e8ca3
 
0a40186
1c8e365
 
0a40186
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1c8e365
 
 
 
 
 
 
 
 
0a40186
 
 
 
 
 
 
87e8ca3
0a40186
 
 
 
87e8ca3
 
 
 
0a40186
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
---
license: mit
tags:
  - interpretability
  - mechanistic-interpretability
  - activation-steering
  - llm
  - kv-cache
  - icml-2026
---

# DecodeShare

Large artifacts for **DecodeShare: Tracing the Shared Pathways of LLM
Decode-Time Decisions**.

The corresponding GitHub release branch is:

```text
https://github.com/Zishan-Shao/decodeshare/tree/Halo
```

This Hugging Face repository is intended for files that should not live in Git
history:

- decode-time activation caches
- shared-subspace bases
- patchback result archives
- downstream compression outputs, with oversized profiling caches stored as
  `.pt.part-*` chunks
- selected steering vectors and cached candidate pools
- rebuttal mechanism and scaling artifacts that are too bulky for the main GitHub
  branch

The GitHub branch tracks compact code, scripts, summaries, and the full artifact
manifest at:

```text
docs/artifact_manifest.tsv
```

Suggested layout:

```text
artifacts/
  Hype1/results/acts/
  patch_back/results/
  downstream/outputs/
  rebuttal/
  results/rebuttal_mechanism/
  results/rebuttal_scaling/
```

The 2026-05-10 rebuttal upload is summarized in:

```text
docs/REBUTTAL_UPLOAD_2026-05-10.md
```

Install the Hugging Face CLI and upload from the original workspace:

```bash
pip install -U huggingface_hub[hf_transfer]
hf auth login
cd /path/to/decodeshare
hf upload Zishan-Shao/decodeshare Hype1/results/acts artifacts/Hype1/results/acts
hf upload Zishan-Shao/decodeshare patch_back/results artifacts/patch_back/results
```

For downstream profiling caches, use the split-file workflow in
`docs/HUGGINGFACE_UPLOAD.md`. Reassembly notes are included under
`artifacts/downstream/outputs/SPLIT_FILES.md` after upload.

Model and dataset licenses remain governed by their upstream providers.