Datasets:
Dataset card
Browse files
README.md
CHANGED
|
@@ -1,3 +1,61 @@
|
|
| 1 |
---
|
| 2 |
license: apache-2.0
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 3 |
---
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
---
|
| 2 |
license: apache-2.0
|
| 3 |
+
tags:
|
| 4 |
+
- magic-the-gathering
|
| 5 |
+
- draft
|
| 6 |
+
- model-evaluation
|
| 7 |
+
- reproducibility
|
| 8 |
+
- predictions
|
| 9 |
---
|
| 10 |
+
|
| 11 |
+
# DraftFM frozen_eval: the per-pick prediction archive
|
| 12 |
+
|
| 13 |
+
Every prediction behind the tables in the DraftFM paper, archived so the
|
| 14 |
+
results can be recomputed (no GPU needed) and audited pick by pick.
|
| 15 |
+
|
| 16 |
+
- **Model:** [brianward92/draftfm](https://huggingface.co/brianward92/draftfm)
|
| 17 |
+
- **Code and paper:** https://github.com/brianward92/mtga
|
| 18 |
+
|
| 19 |
+
Each top-level directory is named by the sha256 of the frozen data snapshot
|
| 20 |
+
it was evaluated against — the directory name itself is the registration.
|
| 21 |
+
Inside are per-pick prediction parquets for every model in the battery
|
| 22 |
+
(scaling rungs `s1`–`s16`, `f-dev`, `f-full`, ablations, and baselines),
|
| 23 |
+
in both `human` and `deployment` scoring modes, plus derived accuracy
|
| 24 |
+
curves and a `summary.json` recording the evaluation context and gates.
|
| 25 |
+
|
| 26 |
+
## Directories
|
| 27 |
+
|
| 28 |
+
| Directory | What it is |
|
| 29 |
+
|---|---|
|
| 30 |
+
| `013df16b…183a` | **The headline MSH pass**: single pre-registered zero-shot evaluation over the first public MSH snapshot (PremierDraft, protocol v1.1, executed 2026-07-07). Source of the 57.0% top-1 result |
|
| 31 |
+
| `013df16b…183a_postday1` | Post-day-one MSH ceiling: per-set model fine-tuned on MSH after the zero-shot pass, as pre-registered in the protocol (2026-07-09) |
|
| 32 |
+
| `df6220d8…853` | Full-battery end-to-end rehearsal on SOS PremierDraft (protocol v1, 2026-07-05), run before the real MSH pass to burn in the pipeline |
|
| 33 |
+
| `6552b298…34e` | Earlier SOS TradDraft rehearsal with baselines (protocol v1, 2026-07-04) |
|
| 34 |
+
| `validation/` | Per-set validation parquets (e.g. SOS) used for sanity gates |
|
| 35 |
+
|
| 36 |
+
The rehearsals are included deliberately: the protocol's credibility comes
|
| 37 |
+
from showing the pipeline was exercised and frozen *before* the one-shot
|
| 38 |
+
MSH evaluation, not tuned after it.
|
| 39 |
+
|
| 40 |
+
## Using it
|
| 41 |
+
|
| 42 |
+
Each `<model>.<mode>.parquet` holds one row per pick with the model's
|
| 43 |
+
probability over the pack and the chosen card; `summary.json` in each
|
| 44 |
+
directory records the snapshot sha256, S3 ETag, protocol tag, execution
|
| 45 |
+
timestamp, and data-quality gates. To recompute a table, aggregate the
|
| 46 |
+
parquets; to audit a claim, join any pick back to its prediction. The
|
| 47 |
+
battery definition lives in the model repo's `frozen_battery.json`; the
|
| 48 |
+
chronology is in the code repo's `docs/eval_protocol.md`.
|
| 49 |
+
|
| 50 |
+
## Provenance and licensing
|
| 51 |
+
|
| 52 |
+
These files are model outputs and derived aggregates generated by the
|
| 53 |
+
DraftFM pipeline (Apache-2.0). The underlying human draft data is the
|
| 54 |
+
17Lands public dataset (CC BY 4.0 — "Data from 17Lands.com"); raw
|
| 55 |
+
17Lands rows are not redistributed here. Unofficial Fan Content per the
|
| 56 |
+
Wizards of the Coast Fan Content Policy; not approved or endorsed by
|
| 57 |
+
Wizards.
|
| 58 |
+
|
| 59 |
+
## Contact
|
| 60 |
+
|
| 61 |
+
Brian Ward — brian.ward.92@gmail.com
|