Juchen/league-toxicity-chat

League of Legends chat toxicity: training and eval data

Built from commit d0ad698 of league-toxicity-slm. The model trained on it is Juchen/league-toxicity-qwen3-1.7b.

Files here

Here	In the GitHub repo	What
sft/	data/sft/	The exact training files, chat format, as trained (hashes in sft/manifest.json)
flat/	(built from sft/)	The same rows as id, message, label, source, category, for the viewer and load_dataset
teacher/	data/sft/teacher_v2*	The 676 kept teacher lines, the 40 the critic disputed, and the generation log
eval/	data/eval/	The test split (FULL) and its first 1,000 rows (E1000), hashes in eval/manifest.json
pressure/	data/pressure/	The 40-line pressure set

from datasets import load_dataset
train = load_dataset("Juchen/league-toxicity-chat", "train_v2")       # 11,595 training rows + 1,213 validation
test = load_dataset("Juchen/league-toxicity-chat", "test", split="e1000")

Every file the model was trained or scored on: its shape, where it came from and who labelled it. Hashes are in data/manifest.json and data/sft/manifest.json.

At a glance

File	Rows	Toxic	Labels from	Used for
data/sft/train_v1.jsonl	10,919	1,006 (9.2%)	L2DTnH experts	Training v1, and the data-efficiency curve (its first N rows)
data/sft/teacher_v2.jsonl	676	269 (39.8%)	The category it was written for, under the written rules	Mixed into v2 and v2-small
data/sft/train_v2.jsonl	11,595	1,275 (11.0%)	10,919 expert + 676 teacher	Training v2, the shipped model
data/sft/train_v2_small.jsonl	3,405	501 (14.7%)	2,729 expert + 676 teacher	Training v2-small
data/sft/train_v3.jsonl	13,623	2,082 (15.3%)	10,919 expert + the same 676 teacher lines, each 4 times	Training v3, a heavier dose of the teacher lines
data/sft/val.jsonl	1,213	112 (9.2%)	L2DTnH experts	Held aside; not used for training or any reported score
data/eval/E1000.jsonl	1,000	102 (10.2%)	L2DTnH experts	The benchmark: every precision, recall and F1 in the reports
data/eval/FULL.jsonl	3,035	280 (9.2%)	L2DTnH experts	The whole test split; confirmed the best prompted cell
data/pressure/PRESSURE.jsonl	40	11 (27.5%)	The project author, under LABELING_RULES.md	Robustness: 4 hard lines for each of 10 kinds
data/sft/teacher_v2_disputed.jsonl	40	—	Teacher and critic disagreed	Left out of training
data/smoke/*	20 teacher + 260 mixed	—	As above	The smoke test only (train/smoke.sh)

Expert-labelled vs teacher-written. v1 trained only on expert-labelled real chat. v2 adds 676 teacher-written lines, 5.8% of its data. v3 repeats each of those lines 4 times, so they are 19.8% of its rows (2,704); a repeat's id ends in -r0 to -r3. Teacher rows have ids starting t2- and a category field; expert rows have the L2DTnH message id. Every eval set is real chat, except the pressure set, which the author wrote and labelled.

Where the real chat comes from

L2DTnH (Ave, Pekaric, Frohner, Apruzzese, 2026) is League of Legends chat from reported games, released under the MIT licence at irdin-pekaric/esorics26_toxicity. This project pins commit 66929d13e455d529f99feee2819374a60ecd6ad9. The raw CSVs are not committed; scripts/get_data.py downloads them at that commit and checks their hashes and row counts.

• Raw files: a train split of 12,136 messages (1,118 toxic) and a test split of 3,035 (280 toxic). Each row has the message, the speaker's relation to the reported player (offender, ally or enemy), a timestamp, report counts and reasons, the chat log id, the champion and the label. Two more columns, old_label and wrongly classified, show a second review pass that the paper doesn't describe. It covered 1,526 messages across both splits and changed 630 labels: 365 toxic to fine and 265 fine to toxic.
• How the experts labelled, per the paper:
  • Eight players with 6 to 20 years of League experience labelled the first 5,000 messages independently. A message was toxic if at least 2 of them flagged it (Fleiss' κ 0.62; 4,704 of the 5,000 were unanimous).
  • Three of them labelled the rest, discussing disagreements until they agreed.
  • They could see the surrounding chat; the model sees one line.
• What this project built from it (scripts/make_sft_data.py, seed 7):
  • Train split: 4 rows with empty messages were dropped. The other 12,132 were split, stratified by label, into train_v1.jsonl (90%) and val.jsonl (10%). train_v1.jsonl is shuffled once, so its first 1,364, 2,729, 5,459 and 10,919 rows are nested subsets.
  • Overlap check: the script asserts that no training id appears in the test split.
  • Test split: FULL.jsonl is the whole test split. E1000.jsonl is its first 1,000 rows (898 fine, 102 toxic), the benchmark used everywhere.

Where the teacher lines come from

scripts/gen_teacher_v2.py asked Claude Opus 5.5 for League chat lines in 12 categories, aimed at the two failures v1 showed on the pressure set. Each category fixes the label, under the written rules:

Toxic categories (269 kept)	Not-toxic categories (407 kept)
sarcasm at a teammate (57), taunting the enemy (48), "diff" put-downs (56), polite put-downs (48), disguised insults (60)	quoting or reporting (59), game complaints (60), self-criticism (59), harmless trigger words (59), feed callouts (60), friendly banter (50), telling a flamer to stop (60)

The gates. Of 720 generated lines:

• None were duplicates or outside 2–120 characters.
• 4 were dropped by the leakage check: 2 matched an eval message and 2 were within 0.8 similarity of a pressure line.
• A GPT-5.6 Sol critic labelled the remaining 716 under LABELING_RULES.md. It disputed 40, which were left out (teacher_v2_disputed.jsonl).

The teacher calls and the critic cost $0.81 at list prices (data/sft/teacher_v2_meta.json). scripts/make_sft_v2.py (seed 11) mixes the kept lines into the expert rows. scripts/make_sft_v3.py (the same seed) mixes in 4 copies of each instead; no new lines were generated.

The author's labels

• Pressure set (data/pressure/PRESSURE.jsonl): 40 lines the author wrote to trip models, 4 for each of 10 kinds, labelled under the written rules: sarcasm, quoting or reporting, self-criticism, talk about the game, harmless trigger words, banter, blunt callouts, taunts, blame without an insult, and disguised insults. No teacher line is within 0.8 similarity of any of them.
• Relabel study (relabel/): 140 messages the author relabelled blind, with the surrounding chat, to see whose standard the prompted models' errors reflect (results/relabel_analysis.md). Not used for training or scoring.

Formats

All files are JSON Lines, UTF-8, with LF line endings (the hashes depend on it).

// Training rows (train_v1, train_v2, train_v2_small, train_v3, val): the exact prompt the model sees
{"id": "15922", "label": 0,
 "prompt": [{"role": "system", "content": "You moderate League of Legends chat. ... Reply with exactly one word: toxic or nontoxic."},
            {"role": "user", "content": "place"}],
 "completion": [{"role": "assistant", "content": "nontoxic"}]}
// Teacher rows also carry "category"; their ids start with "t2-".

// Eval rows (E1000, FULL)
{"id": "9524", "message": "this mid", "label": 0, "chatlog_id": "1005.0", "time": "00:16:56", "champion_name": "Jarvan IV"}

// Pressure rows
{"id": "p-01", "message": "wow great gank jungle, really helpful", "label": 1, "kind": "sarcasm at a teammate"}

label is 1 for toxic and 0 for not toxic throughout.

Known limits

• The two label sources disagree on some lines. The experts labelled many bare feed callouts toxic, and borderline lines lean toxic under the 2-of-8 rule. The written rules lean not toxic when context is missing. See LABELING_RULES.md.
• One line, no context. The experts saw the chat around each message; the model doesn't. Some expert labels can't be recovered from the line alone (GAREN TANK, 19 assist).
• Toxic lines are rare: about 9% of real chat. Recall on 102 toxic benchmark lines has wide intervals.
• The teacher lines come from one model. Their idea of sarcasm or a taunt is Claude's, which may differ from how players write.
• English only, and only from reported games. The most common report reasons are verbal abuse, negative attitude, offensive language and assisting the enemy team, so the chat skews heated.
• These are real players' messages. Use them for research only, and never to label a real person as toxic.