Rename configs ace->trigger_and_args, distant->trigger_only (task-based naming); add provenance/license notes

.gitattributes

@@ -58,4 +58,4 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 # Video files - compressed
 *.mp4 filter=lfs diff=lfs merge=lfs -text
 *.webm filter=lfs diff=lfs merge=lfs -text
-data/fewevent_distant_test.jsonl filter=lfs diff=lfs merge=lfs -text
+data/fewevent_trigger_only_test.jsonl filter=lfs diff=lfs merge=lfs -text

README.md

@@ -11,15 +11,15 @@ features:
       - name: description
         dtype: string
 configs:
-  - config_name: ace
+  - config_name: trigger_and_args
     default: true
     data_files:
       - split: test
-        path: data/fewevent_ace_test.jsonl
-  - config_name: distant
+        path: data/fewevent_trigger_and_args_test.jsonl
+  - config_name: trigger_only
     data_files:
       - split: test
-        path: data/fewevent_distant_test.jsonl
+        path: data/fewevent_trigger_only_test.jsonl
 ---
 
 # FewEvent (mneb/casie format)
@@ -28,16 +28,36 @@ FewEvent (Deng et al., WSDM 2020) converted into the mneb `json_structures` even
 format. FewEvent annotates two tasks — event **detection** (triggers) and event **argument
 extraction** (roles); it has no entity or relation layer, so each record's `output` carries
 only `json_structures`. Char offsets are end-exclusive (`input[start:end] == text`).
+One record = one sentence = one event. Single `test` split per config.
 
-The source mixes two annotation provenances at the instance level, kept here as **two
-separate configs**:
+## The two configs (split by what annotation is present)
 
-- **`ace`** (default) — ~16.4k instances derived from ACE/KBP, with clean trigger **and
-  typed-argument** spans (33 event types, 35 ACE roles). Both tasks are present.
-- **`distant`** — ~65k Freebase/Wikipedia distant-supervised instances (67 further event
-  types) that carry a usable trigger but no clean role-typed arguments, so they are shipped
-  **trigger-only** (`arguments: []`).
+- **`trigger_and_args`** (default) — ~16.4k records that carry a clean trigger **and typed
+  arguments** (33 event types, 35 roles). Supports both event detection and argument
+  extraction.
+- **`trigger_only`** — ~65k records with a trigger but **no arguments** (`arguments: []`).
+  Supports event detection only.
 
-One record = one sentence = one event. Single `test` split per config. See
-`fewevent_label_summary.md` for the full type/role inventory and counts, and
+The split criterion is the annotation itself (does the instance have clean role-typed
+arguments?), i.e. which task it supports — **not** the raw data source. Provenance correlates
+but does not line up cleanly:
+
+- `trigger_and_args` is essentially the **human-annotated** portion (ACE-2005 + TAC-KBP-2017).
+- `trigger_only` is **mostly** the Freebase/Wikipedia **distant-supervised** portion, but also
+  includes a few thousand ACE/KBP-origin instances that simply lack clean argument
+  annotations (the raw enriched file was augmented unevenly). So these are named by task, not
+  by source.
+
+## Provenance & license (please read before redistributing)
+
+FewEvent was built from: **ACE-2005** (LDC2006T06) + **TAC-KBP-2017 Event Track** (both
+human-annotated, **LDC-licensed / restricted**), and **Freebase + Wikipedia** auto-labeled
+data (distant supervision; Wikipedia text is CC-BY-SA, Freebase was CC-BY). The upstream
+GitHub release (`231sm/Low_Resource_KBP`) ships **no license file**. The `trigger_and_args`
+half therefore contains LDC-derived text whose public redistribution may be restricted;
+downstream users should ensure they hold the appropriate ACE-2005 / TAC-KBP licenses. The
+`trigger_only` half's distant-supervised labels are **noisy** — treat as eval-only, and do
+not treat absence of an event as ground truth.
+
+See `fewevent_label_summary.md` for the full type/role inventory and counts, and
 `fewevent_vis.html` for a sampled visualiser.

fewevent_label_summary.md

@@ -1,44 +1,61 @@
 # FewEvent — label summary
 
-Counts computed from `processed_data/fewevent/fewevent_ace_test.jsonl` (16,364 records) and
-`processed_data/fewevent/fewevent_distant_test.jsonl` (65,093 records), produced by
-`scripts/preprocess_fewevent.py` from the **enriched** `Few-Shot_ED.json` (85,569 instances,
-git clone `231sm/Low_Resource_KBP`).
+Counts computed from `processed_data/fewevent/fewevent_trigger_and_args_test.jsonl` (16,364
+records) and `processed_data/fewevent/fewevent_trigger_only_test.jsonl` (65,093 records),
+produced by `scripts/preprocess_fewevent.py` from the **enriched** `Few-Shot_ED.json` (85,569
+instances, git clone `231sm/Low_Resource_KBP`).
 
 FewEvent annotates two tasks — event **detection** (triggers) and **argument extraction**
 (roles) — and has **no entity layer and no relations**, so `output` carries only
 `json_structures`. One record = one sentence = one event; single `test` split per group.
 
-## Two provenance groups (kept separate by request)
+## Two configs (split by what annotation is present)
 
-The raw file mixes two annotation qualities **per instance** (a single event type can hold
-both), split here into two files / two HF configs:
+The split criterion is the annotation itself — **does the instance carry clean role-typed
+arguments?** — i.e. which TASK it supports. It is **not** the raw data source (see provenance
+note below).
 
-| Group (config) | Records | Event types | Tasks present | Quality |
-|---|---:|---|---|---|
-| **ace** (default) | 16,364 | 33 | trigger **+ typed arguments** (35 roles) | ACE/KBP-derived, clean char spans |
-| **distant** | 65,093 | 79 | trigger only (`arguments: []`) | Freebase/Wikipedia distant-supervised |
+| Config | Records | Event types | Tasks | What's in a record |
+|---|---:|---:|---|---|
+| **trigger_and_args** (default) | 16,364 | 33 | ED + EAE | trigger **+ typed arguments** (35 roles) |
+| **trigger_only** | 65,093 | 79 | ED only | trigger, `arguments: []` |
 
-The two groups' type sets overlap: 21 types are ace-only, 67 are distant-only, and **12
-appear in both** (e.g. `Conflict.Attack` = 4407 ace + 354 distant, `Life.Marry` = 205 ace +
-25170 distant) — same type, different-provenance instances. Union = **100 event types**.
+The two groups' type sets overlap: 21 types are trigger_and_args-only, 67 are
+trigger_only-only, and **12 appear in both** (e.g. `Conflict.Attack` = 4407 + 354,
+`Life.Marry` = 205 + 25170). Union = **100 event types**.
 
-Scope caveats: (1) the `distant` group is distant-supervised (noisy triggers, no gold args) —
-use as such; (2) `arguments` exist only on `ace`; (3) event types kept **verbatim**
-(`Major.Subtype`, dot form), roles verbatim; (4) FewEvent is a few-shot corpus with no native
-train/dev/test — everything is one combined `test` split (TextEE's 5 random re-splits are not
-used). Use as **eval-only**.
+### Provenance note (why task-based names, not `ace`/`distant`)
+
+Provenance correlates with the split but does **not** line up cleanly:
+
+- **trigger_and_args** ≈ the human-annotated **ACE-2005 + TAC-KBP-2017** portion (LDC-licensed,
+  restricted).
+- **trigger_only** is **mostly** the **Freebase/Wikipedia distant-supervised** portion
+  (noisy; Wikipedia CC-BY-SA, Freebase CC-BY), but also includes a few thousand ACE/KBP-origin
+  instances that simply lack clean argument annotations (the enriched file was augmented
+  unevenly). Naming these `distant` would over-claim the source, so we name by the task each
+  supports.
+
+Structurally, the raw non-clean-args instances come in three forms, all emitted trigger-only:
+Freebase MID+relation block (~63.6k), a 4th element that is just a re-cleaned sentence
+(~4.4k, mostly ACE-domain types), and bare 3-element tuples (~1.2k).
+
+Scope caveats: (1) the `trigger_only` labels are distant-supervised (noisy) — use as such;
+(2) event types kept **verbatim** (`Major.Subtype`, dot form), roles verbatim; (3) FewEvent
+has no native train/dev/test — everything is one combined `test` split (TextEE's 5 random
+re-splits are not used). Use as **eval-only**.
 
 Build notes: doc_ids assigned by TextEE's `sorted(types)` enumeration (to key into
-`token_map.json`, the trigger-span source for the `distant` group). `distant` dropped 4,112
-instances with no token_map entry (4,102) or >300 tokens (10). Trigger/arg text is the
-authoritative `input[start:end]` slice; it mismatched the provided gold word in 101 `ace`
-spans (~0.6%, mostly hyphen sub-token artifacts like `near-collapse`→`collapse`) and 1
-`distant` span — offsets remain valid (0 offset-invariant failures).
+`token_map.json`, the trigger-span source for the `trigger_only` group). `trigger_only`
+dropped 4,112 instances with no token_map entry (4,102) or >300 tokens (10). Trigger/arg text
+is the authoritative `input[start:end]` slice; it mismatched the provided gold word in 101
+`trigger_and_args` spans (~0.6%, mostly hyphen sub-token artifacts like
+`near-collapse`→`collapse`) and 1 `trigger_only` span — offsets remain valid (0
+offset-invariant failures).
 
 ## Totals at a glance
 
-| Item | ace | distant |
+| Item | trigger_and_args | trigger_only |
 |---|---:|---:|
 | Records (sentences / events) | 16,364 | 65,093 |
 | Event types | 33 | 79 |
@@ -47,7 +64,7 @@ spans (~0.6%, mostly hyphen sub-token artifacts like `near-collapse`→`collapse
 
 Combined records: **81,457**. Union event types: **100**.
 
-## ace — event types (33, 16,364 events)
+## trigger_and_args — event types (33, 16,364 events)
 
 | Event type | Count |
 |---|---:|
@@ -86,7 +103,7 @@ Combined records: **81,457**. Union event types: **100**.
 | Justice.Pardon | 10 |
 | **Total** | **16364** |
 
-## ace — argument roles (35, 33,170 mentions)
+## trigger_and_args — argument roles (35, 33,170 mentions)
 
 | Role | Count |
 |---|---:|
@@ -127,7 +144,7 @@ Combined records: **81,457**. Union event types: **100**.
 | Price | 32 |
 | **Total** | **33170** |
 
-## distant — event types (79, 65,093 events, trigger-only)
+## trigger_only — event types (79, 65,093 events, trigger-only)
 
 | Event type | Count |
 |---|---:|