osieosie commited on
Commit
0c4e432
·
verified ·
1 Parent(s): 50e1356

Add README: document the four-directory plain-image layout

Browse files
Files changed (1) hide show
  1. README.md +92 -0
README.md ADDED
@@ -0,0 +1,92 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ---
2
+ license: other
3
+ tags:
4
+ - terminal-agent
5
+ - reinforcement-learning
6
+ - apptainer
7
+ ---
8
+
9
+ # tmax image pool
10
+
11
+ Every Apptainer/SIF image referenced by the three `swerl-tmax-15k` variants, stored once
12
+ and shared between them.
13
+
14
+ ## Read the layout from the repo, not from a prefix
15
+
16
+ **The plain images are split across four directories, not one.** A consumer that filters
17
+ on `images/` silently gets 4,519 fewer files and then reports itself complete — the
18
+ tasks that reference the missing blobs fail later, at sandbox start, one by one.
19
+
20
+ | directory | files |
21
+ |---|---:|
22
+ | `images/` | 9,971 |
23
+ | `images_b00/` | 1,933 |
24
+ | `images_b01/` | 2,000 |
25
+ | `images_b02/` | 586 |
26
+ | **plain, total** | **14,490** |
27
+ | `derived/` | 1,713 |
28
+ | **total SIFs** | **16,203** |
29
+
30
+ The batching exists only because HuggingFace enforces a hard limit of 10,000 files per
31
+ directory at push time; `images_b00..b02` are continuations of `images/` with no other
32
+ meaning. More batches may be added, so **enumerate whatever directories exist** rather
33
+ than hard-coding this list.
34
+
35
+ Resolve paths through `task_to_blob.json` instead of guessing — it already carries the
36
+ correct directory for every task.
37
+
38
+ ## `task_to_blob.json`
39
+
40
+ Keyed by variant, then by task id, valued by the blob path within this repo:
41
+
42
+ ```json
43
+ {
44
+ "original": {"task_000868_f0470dff": "images_b00/hamishi740__swerl-tmax-v3__17c682d5c609.sif", ...},
45
+ "hardened_prefilter": {...},
46
+ "solvable": {...}
47
+ }
48
+ ```
49
+
50
+ | variant | tasks | unresolvable |
51
+ |---|---:|---:|
52
+ | `original` — [`hamishivi/swerl-tmax-15k`](https://huggingface.co/datasets/hamishivi/swerl-tmax-15k) | 14,601 | 0 |
53
+ | `hardened_prefilter` — [`wAI-org/swerl-tmax-15k-hardened-prefilter`](https://huggingface.co/datasets/wAI-org/swerl-tmax-15k-hardened-prefilter) | 13,565 | 0 |
54
+ | `solvable` — [`wAI-org/swerl-tmax-15k-solvable-gpt-5-6-terra`](https://huggingface.co/datasets/wAI-org/swerl-tmax-15k-solvable-gpt-5-6-terra) | 7,015 | 0 |
55
+
56
+ Tasks outnumber blobs because distinct tasks share an image.
57
+
58
+ ## Plain vs derived
59
+
60
+ `images*/` are the registry images unchanged — a fresh `apptainer pull` of the original
61
+ reference is byte-identical to the stored SIF.
62
+
63
+ `derived/` are single-step derivations used by the hardened variants: an oracle file
64
+ removed, and/or the dependency the verifier imports installed, with per-task pinned
65
+ versions. A task in a hardened variant may map to either kind, so do not assume a
66
+ variant uses only one directory.
67
+
68
+ ## Using it
69
+
70
+ ```python
71
+ import json
72
+ from huggingface_hub import hf_hub_download, snapshot_download
73
+
74
+ m = json.load(open(hf_hub_download("wAI-org/tmax-image-pool", "task_to_blob.json",
75
+ repo_type="dataset")))
76
+ blobs = sorted(set(m["solvable"].values())) # only what this variant needs
77
+ snapshot_download("wAI-org/tmax-image-pool", repo_type="dataset",
78
+ allow_patterns=blobs, local_dir="sifs/")
79
+ ```
80
+
81
+ Then point each row's `env_config.image` at the absolute local `.sif` path;
82
+ `prefer_local_sif()` passes an absolute path straight through with no registry lookup.
83
+
84
+ Fetching only one variant's blobs is much cheaper than the whole pool — `solvable` needs
85
+ 6,968 of the 16,203.
86
+
87
+ ## Verifying what you fetched
88
+
89
+ Check `apptainer inspect` on each distinct SIF a variant references. Size and hash
90
+ agreement is not sufficient: SIFs have been seen to match both and still not be valid
91
+ squashfs. Count coverage per variant against `task_to_blob.json` rather than trusting a
92
+ fetcher's own "done".