praxis-chess/praxis-chess-evidence-graphs

Praxis Chess Evidence Graphs

Built for, and used to train and test, praxis-chess-reasoner-qwen3.5-2b-lora and its 4B sibling.

Chess mistakes from public Lichess data, each with an evidence graph of engine-computed facts about the move, and a verified explanation: a typed reasoning chain that a program can check claim by claim, the labels (what happened, and why) and two or three sentences of prose.

The dataset exists to test one question: does a small model explain mistakes more truthfully when it is given structured evidence? Every row is rendered four ways, from the bare board (R0) to the full graph (R3), and the same target answer is paired with all four, so the input is the only thing that changes between the training sets.

What is in it

Path	Rows	What
data/R0 … data/R3 / train.jsonl	7,057	Training rows, chat format, at each rendering
data/R0 … data/R3 / validation.jsonl	1,121	Validation rows, uncapped (checkpoint selection used the first 200)
data/graphs/train.jsonl, validation.jsonl	7,057 / 1,121	The evidence graph and the rules' labels behind each training and validation row, by id (= meta.source_id)
data/test/test.jsonl	838	The public test set: graph, every rendering, the rules' labels
results/outputs.jsonl	5,028	Every model answer behind the published table, as generated
results/verified.jsonl	5,866	Those answers with the checker's verdict, plus the rules' own diagnosis of each item
results/public_test_v1.json, .md		The published table, with 95% intervals
schema/graph_schema.json		JSON Schema of the evidence graph (version 2)
schema/answer_schema.json		The answer's JSON schema, claim types and their required arguments
verifier/verifier_test_vectors.jsonl.gz	1,959	Graph + answer + the Java checker's verdict, for porting the checker
manifest.json		Counts, provenance, and the sha256 of every file

The four renderings

The user message a model sees. Each level contains the one below it.

• R0: the position (FEN), the player's side and the move played.
• R1: R0 plus the engine's best move, the evaluation before and after, and the engine line after the played move.
• R2: R1 plus a flat list of what the move changed: what it attacks, pieces left loose or without a defender, lines opened or closed.
• R3: the evidence graph: everything above plus the opponent's threat before the move (a null-move probe), the counterfactual (does the punishing reply still work after the best move?) and how deep the engine had to look, as items with IDs (P0, R1, CF1, T1, Δ1, D1, …) that a claim can cite.

A training row

{"messages": [{"role": "system", "content": "You are Prax. Diagnose the mistake using ONLY the evidence given. Every claim must be typed. Output JSON."},
              {"role": "user", "content": "<the rendering>"},
              {"role": "assistant", "content": "<the target answer, JSON>"}],
 "meta": {"slice": "A", "source_id": "lichess:puzzle:…", "R": "R3", "consequence": "LOST_MATERIAL",
          "mechanism": "IGNORED_THREAT", "rules_fired": ["IGNORED_THREAT"], "composite": false,
          "teacher": "…", "verified": true, "tokens_in": 471, "…": "…"}}

The answer is JSON: reasoning_chain (at most 7 typed claims, each with type, args, cites and text), then consequence, mechanism, motif, critical_response, visibility and explanation. The claim types and their arguments are in schema/answer_schema.json.

A test item

id, slice, severity, fen_key, graph (the full evidence graph), renders (R0–R3 plus the four ablated R3 renderings used in the study: R3-CF, R3-T, R3-DELTA, R3-D, each with one component removed, and a token count for each) and rules (the rules' labels: consequence, mechanism, which rules fired, single-cause or composite, motif, visibility).

How it was built

1. Sources (CC0):

   • Slice A, puzzles: 3,049 positions from the Lichess puzzle database, chosen with per-mechanism quotas.
   • Slice B, games: 5,967 mistakes from rated rapid and blitz games between players rated 1000–1600, from the June 2024 Lichess export. They were found by Praxis's own analysis pipeline, as the app would flag them. Player names, ratings, dates and links were removed before analysis. Each game is known only by a hash of its URL.

2. Evidence: for each mistake, Praxis's EvidenceGraphBuilder runs Stockfish (deterministic: one thread, hash cleared, depth 14) and records typed facts, never labels. Graph version 2; backend commit aae5ea9bb291.

3. Labels and targets: Praxis's rules label every mistake (what happened, which mechanism rules fired) and write a verified diagnosis wherever exactly one rule fires. A teacher model, Qwen/Qwen3.5-27B-FP8, then:

   • rewrote the prose: 3,968 of 5,746 attempts kept;
   • answered the mistakes with more than one possible cause: 2,102 of 3,496 attempts passed, and 1,703 entered training and validation after the caps.

   Only answers that passed the checker were kept. 5,087 of the 7,057 training targets contain teacher-written text; the rest are entirely the rules' own.

4. Clean-up and split:

   • one row per position;
   • any position that also appears among the author's private test mistakes or the few-shot examples was dropped (53 rows);
   • the split is by source, so a puzzle or a whole game lands in exactly one of train, validation and test;
   • training rows are capped per mechanism and per slice, and "nothing concrete" rows are capped at a share, so no single answer dominates;
   • the test set is not capped.

Split	Rows	Slice A	Slice B	Lost material	Missed material	Mated	Missed mate	Nothing concrete	Composite
Train	7,057	2,549	4,508	4,512	696	426	53	1,370	1,575
Validation	1,121	264	857	464	151	60	15	431	128
Test	838	236	602	350	102	46	3	337	112

Using it

from datasets import load_dataset

train = load_dataset("praxis-chess/praxis-chess-evidence-graphs", "R3", split="train", revision="v1.0")
test = load_dataset("praxis-chess/praxis-chess-evidence-graphs", "test", split="test", revision="v1.0")

Checking answers. praxis_eval (Python, needs only python-chess) is a port of the Java checker that judged every published number:

• it matches the Java verdict on all 5,866 published answers and on the 1,959 test vectors here, broken answers included;
• claim mode checks every claim against the full graph and the board;
• citation mode also requires each cited ID to exist and to be about the claim's squares.

from praxis_eval import parse_diagnosis, verify

report = verify(item["graph"], parse_diagnosis(model_answer))
report.passed, report.claims_true, report.claims

Reproducing the table. In the Praxis repo:

• python reproduce.py score re-checks the published answers here;
• python reproduce.py run --model 2b-r3 regenerates them.

See REPRODUCE.md.

Results on the test set

System	Items	Parsed	Claim precision (95% CI)	Chain valid (95% CI)	Mechanism vs rules	Invents a cause	Median latency
Rules (no model)	838	100%	100.0% (100.0–100.0)	100.0% (100.0–100.0)	100%	0%	—
2B LoRA, R0 (board only)	838	100%	34.8% (31.4–38.1)	0.5% (0.1–1.1)	1%	1%	4.0 s
2B LoRA, R1 (+ engine facts)	838	100%	56.9% (54.7–58.9)	14.4% (12.1–16.8)	50%	26%	6.6 s
2B LoRA, R2 (+ tactics)	838	100%	56.8% (54.6–58.8)	12.6% (10.6–14.9)	56%	27%	5.9 s
2B LoRA, R3 (full evidence graph)	838	100%	99.7% (99.6–99.9)	78.0% (75.4–80.8)	79%	12%	5.7 s
4B LoRA, R0 (board only)	838	100%	32.6% (29.5–35.8)	0.6% (0.1–1.2)	0%	0%	7.9 s
4B LoRA, R3 (full evidence graph)	838	100%	99.7% (99.5–99.9)	86.8% (84.4–89.0)	95%	16%	16.1 s

Privacy

Every row comes from public Lichess data. Slice B games were scrubbed before analysis: no names, ratings, dates or links. The author's own games, which are the study's primary test set, are not in this dataset in any form. Positions they share with it were removed from training, as described above.

Limitations

• Tactics and material only. Version 1 explains mistakes where material or mate changes hands within 8 plies; the rest are labelled "nothing concrete" with no cause.
• The labels are the rules'. They are hand-written, and they were checked against human labels only on the author's games.
• The teacher's prose is checked, not proofread. The checker guarantees that every move and square it names was earned by a true claim. It does not guarantee good style.
• Narrow source. Lichess puzzles, and 1000–1600 rapid and blitz games. English only.

Licence

• This dataset: CC-BY-4.0, built from CC0 Lichess data.
• Teacher answers: written by Qwen3.5-27B-FP8 (Apache-2.0).
• Engine: the evidence comes from Stockfish (GPL-3.0), run as a separate program.
• Puzzle themes: the CC0 Lichess export carries theme tags produced by the AGPL lichess-puzzler. That tool is not part of this dataset.

Citation

@misc{praxis_lora_v1_2026,
  title  = {Does structured evidence let a small model explain chess mistakes truthfully?},
  author = {Anand, Tanmay},
  year   = {2026},
  note   = {Praxis Chess, LoRA v1 research write-up},
  url    = {https://github.com/Praxis-Chess/Praxis-Chess}
}