zer0contextlost commited on
Commit
7eb8bca
·
verified ·
1 Parent(s): 66e6b34
Files changed (1) hide show
  1. README.md +17 -5
README.md CHANGED
@@ -34,17 +34,29 @@ Built 2026-11-01. Rebuilt nightly while the pipeline runs.
34
 
35
  | | count |
36
  |---|---|
37
- | Structures generated | 5402 |
38
  | Kept as "discoveries" (stable by ML and not found in reference databases at the time) | 1593 |
39
  | ML hull check finished | 968 |
40
  | ML hull check at or below 0 eV/atom | 0 |
41
- | DFT jobs (passes / fails / still running or queued) | 8 (1 / 2 / 5) |
42
 
43
- ML verdicts: unstable 1979, rock solid 1325, plausible 1213, borderline 722, falls apart 163.
44
- Novelty at generation time: skipped 2864, new composition 967, known 940, new structure 626, rediscovered 5.
 
 
 
 
 
 
 
 
 
 
 
45
 
46
  ### Files
47
- - `structures.jsonl.gz`: one JSON object per structure: formula, space group, CIF, ML energies and verdict,
 
48
  novelty check (with the matching database IDs when it was already known), ML hull result, DFT summary.
49
  - `structures.csv`: the same without CIFs, for spreadsheets.
50
  - `dft_results.csv`: one row per DFT job.
 
34
 
35
  | | count |
36
  |---|---|
37
+ | Structures generated | 5404 |
38
  | Kept as "discoveries" (stable by ML and not found in reference databases at the time) | 1593 |
39
  | ML hull check finished | 968 |
40
  | ML hull check at or below 0 eV/atom | 0 |
41
+ | DFT jobs (passes / fails / still running or queued) | 9 (1 / 2 / 6) |
42
 
43
+ ### How the prompts were chosen, and how each way does
44
+ On 2026-11-01 the autopilot switched from random formulas to mostly substitution-based prompts (see METHODS.md).
45
+ This table compares the sources; "within 0.05" counts ML hull results at most 0.05 eV/atom above the hull.
46
+ New rows need a few nights of hull checks before their numbers mean much.
47
+
48
+ | prompt mode | structures | new | ML hull checked | median e_hull (eV/atom) | within 0.05 | DFT passes |
49
+ |---|---|---|---|---|---|---|
50
+ | random formula (before 11-01) | 3713 | 1327 | 873 | 0.356 | 1 | 0 |
51
+ | surprise (before 11-01) | 1681 | 266 | 95 | 0.133 | 14 | 1 |
52
+ | human | 10 | 0 | 0 | - | 0 | 0 |
53
+
54
+ ML verdicts: unstable 1979, rock solid 1325, plausible 1214, borderline 723, falls apart 163.
55
+ Novelty at generation time: skipped 2865, new composition 967, known 941, new structure 626, rediscovered 5.
56
 
57
  ### Files
58
+ - `structures.jsonl.gz`: one JSON object per structure: formula, space group, CIF, how the prompt was chosen
59
+ (`prompt_mode`, from 2026-11-01), ML energies and verdict,
60
  novelty check (with the matching database IDs when it was already known), ML hull result, DFT summary.
61
  - `structures.csv`: the same without CIFs, for spreadsheets.
62
  - `dft_results.csv`: one row per DFT job.