chayuto/klondike-llm-decisions

Klondike Solitaire LLM Advisor Decisions

Per-decision traces from large language models acting as advisors in Klondike Solitaire, collected to support distillation research and the study of LLM failure modes in sequential decision tasks. Every row records one advisor call against a reproducible game state.

Configs at a glance

Several subsets under one dataset path. Pick the one that fits your use-case; researchers who want everything should use the default. Each config is a set of zstd-compressed Parquet shards, read transparently by datasets; new sessions arrive as additional shards.

Config	Rows	Schema	Best for
client_v1_full_corpus_raw (default)	28904	full interaction (prompt + rawResponse + decision blob + call metadata)	failure-mode research, replay, end-to-end audit
client_v1_teacher_clean_raw	12021	full interaction	fine-tuning, honest training-quality subset (single teacher model, current schema, non-stalled)
client_v1_teacher_clean_lean	12021	derived per-decision (flat schema; see Fields)	quick analytics, lightweight loading, headline-statistics work
client_v1_26b_raw	1949	full interaction	comparison: the gemma-4-26b-a4b-it MoE cohort alone (current schema, stalled decisions kept, 1 winning session)
client_v1_26b_lean	1949	derived per-decision (flat schema)	comparison analytics: the MoE on matched game states

from datasets import load_dataset

# Default -- the full corpus, including failure modes
full = load_dataset("chayuto/klondike-llm-decisions")  # 28904 rows

# The training-friendly subset (filtered, single teacher)
clean_raw  = load_dataset("chayuto/klondike-llm-decisions", "client_v1_teacher_clean_raw")   # 12021 rows
clean_lean = load_dataset("chayuto/klondike-llm-decisions", "client_v1_teacher_clean_lean")  # 12021 rows, flat schema

# The 26B MoE cohort on its own, for comparison
moe_26b    = load_dataset("chayuto/klondike-llm-decisions", "client_v1_26b_raw")  # 1949 rows

The client_v1_26b_* configs are a behavioural-comparison cohort: all 1949 current-schema decisions from the gemma-4-26b-a4b-it MoE, with stalled/loop decisions deliberately kept (1221 of 1949 sit inside a stall, where the teacher_clean configs drop such rows). The corpus now holds 1 winning session for the 26B MoE; the cohort still centres on failure-mode contrast (stalled/loop decisions kept), so treat it as a behavioural contrast set more than as additional training data.

Filtering by model

Every row in *_raw configs carries a model field (e.g. "gemma-4-31b-it", "gemini-3.1-flash-lite"). Use the standard HF .filter() to subset:

ds = load_dataset("chayuto/klondike-llm-decisions")  # full corpus
teacher_only = ds["train"].filter(lambda r: r["model"] == "gemma-4-31b-it")
other_only   = ds["train"].filter(lambda r: r["model"] != "gemma-4-31b-it")

The client_v1_teacher_clean_* configs are already filtered to a single teacher model (currently gemma-4-31b-it); use them if you want a homogeneous training subset without writing a filter. The gemma-4-26b-a4b-it MoE subset is likewise available directly as the client_v1_26b_* configs, no filter needed.

Collection method (client_v1_*)

Collected via an external client-side harness (closed-source) running the Klondike app and capturing every teacher-advisor call. Each game seeds a reproducible deal. Rows are deduplicated by their UUIDv7 id across re-exports; nothing is discarded.

• Collection window: 2026-05-17 to 2026-06-26
• Sessions: 161 distinct game sessions
• Models: gemma-4-31b-it (23235), gemini-3.1-flash-lite (2692), gemma-4-26b-a4b-it (1949), MiniMax-M3 (1028)
• Schema tiers: current (28533), legacy (371)

Planned: server_v1_* configs

A second collection method is being prepared using the open-source MCP server in this project's parent codebase (solitaire_analytics.mcp_server). Server-collected rows will ship as server_v1_* configs under this same dataset path. They will carry agent identity (agent_id, model, provider, app_commit) stamped per decision, plus an infoLevel block per session so perfect-vs-imperfect-information runs are unambiguous. Not yet published.

Fields

*_raw configs

Verbatim interaction records as captured by the harness.

• id: globally unique UUIDv7 for the interaction
• sessionId, turnIndex: game session and move number (current-schema rows)
• model, provider: the advisor model
• prompt: full prompt: Klondike rules + board state JSON + legal-move list
• rawResponse: the model's raw text reply
• decision: parsed moveIndex, confidence, alternativeMoveIndex, boardAnalysis, reasoning
• outcome, token counts, timing: call metadata

*_lean config

Derived per-decision rows, flattened. Built by joining each successful interaction against its parsed prompt + decision.

• id, sessionId, turnIndex, timestamp, model, provider, appCommit
• chosenMoveType, chosenMoveDescribe, moveIndex, nLegalMoves
• confidence, alternativeMoveIndex
• completionProgress, moveCount, perceivedDifficulty: from the prompt metrics block
• foundationCards, faceDownTotal, progressScore, turnsSinceProgress: computed by the ingest from board state
• boardAnalysis, reasoning, thinkingText: agent's natural-language fields

Chosen-move distribution (full corpus)

Move type	Count	Share
draw_card	13545	47%
tableau_to_tableau	8874	31%
tableau_to_foundation	2642	9%
discard_to_tableau	1553	5%
discard_to_foundation	1159	4%
recycle_stock	1131	4%

Failure modes are a feature of *_full_corpus_raw, not a bug

The full corpus deliberately includes sessions where the teacher fails to make progress. These are research signal, not noise. The cleaned configs (*_teacher_clean_*) filter them out per a stall heuristic; the full corpus keeps them so you can study the failure modes directly.

Two documented pathologies recur in the corpus:

1. Doom-loop / oscillation. The teacher rationalises a two-card shuffle (e.g. moving 5C/4D back and forth between two columns) as a 'setup move' even when recentMoves clearly shows the exact reversal was just played. Confidence stays saturated at 0.9+ throughout. Example: session with the longest plateau in the corpus carries 75 consecutive turns of foundation/face-down unchanged.

2. Honest hunt that degrades. Some sessions begin with draw-dominated card hunting (correct behaviour when needed cards are still hidden) and only descend into oscillation after extended no-progress windows. A plateau-only stall detector would over-fire on these; a shuffle-fraction gate is needed alongside the plateau gate to discriminate honest hunt from doom-loop.

Use the progressScore / turnsSinceProgress columns in the *_lean config to locate stalled stretches; use the chosenMoveType distribution within those stretches to classify them.

Known limitations

• Confidence is miscalibrated. Reported confidence spans 0.50 to 1.00 (mean 0.89); the teacher signals near-certainty regardless of board state. Do not treat it as a calibrated probability; in our experience using it as a training-time signal teaches student models to be overconfident.
• Mixed schema versions in client_v1_full_corpus_raw. Older rows lack sessionId / turnIndex / appCommit. Filter on field presence if you need a homogeneous subset, or use the client_v1_teacher_clean_* configs which exclude legacy schema rows.
• Outcome skew. Most logged games were lost or stalled; winning play is under-represented. End-game (foundation_cards > ~10) is particularly sparse. Student models trained on this corpus will lack guidance for late-game transitions.
• Mixed information modes. A few early sessions had perfect-information game state exposed to the advisor; most run under imperfect information. The client_v1_teacher_clean_* configs select a single information mode.
• Move-type skew toward draw_card. Draws are ~50 to 66% of eligible rows in the cleaned configs, reflecting the teacher's tendency to keep drawing when no productive tableau move is obvious. Apply your own re-weighting if this matters for your task.

Build on top of this

This corpus is intentionally public so others can study or build on Gemma 4 31B's Klondike behaviour without reproducing the harvest infrastructure from scratch. The license is permissive (CC-BY-4.0); attribution is the only ask. Some specific ways the data is set up to be useful:

Replay any seed in your browser

Every row carries a sessionId (and most carry a seed derivable from the source repo's data/index/manifest.jsonl). The harvester web UI at https://solitaire.chayuto.com/?seed=<seed> deals the same board deterministically: you can load any seed from this corpus and play or feed it to your own model, then compare your model's decisions against the rows here turn for turn.

Run your own kill-or-continue analysis

The source repo at chayuto/solitaire-analytics publishes the tooling used to produce this corpus, including:

• scripts/ingest_exports.py: the dedup + stall-filter pipeline that produced these configs from raw exports.
• .claude/skills/solitaire-analyst/: a Claude Code skill that reads any raw export and produces a kill-or-continue verdict with failure-mode classification. Includes a Monte Carlo solvability check via pyksolve (DFS with dominance pruning, ~10 ms per sample) at .claude/skills/solitaire-analyst/scripts/check_winnability.py.
• data/DATASET_NOTES.md: the long-form taxonomy of every documented session in the corpus. Each entry calls out the failure class (behavioural-doom-loop, dead-deal-flailing, honest-hunt, self-rescue-fails) with the specific evidence that drove the call. Useful if you want to know which sessions are which kind of failure before pulling them.

Compare a model on the same boards

A 20-state Klondike-state benchmark used by this project's distillation evaluations lives in the source repo under experiments/a4_phase1.5_2026_05_24/prompts/C0/. Five early-game, eight midgame, seven oscillation-prone states; each state's reference answer is the teacher model's pick scored on a six-level tier (foundation > reveal > waste_play > shuffle > draw > illegal). If you want to bench your own Klondike-playing model on the same positions and compare apples-to-apples against gemma-4-31b-it, this is the fastest way.

Cite if you publish

If this corpus shows up in a paper, blog post, or model card, please cite the dataset URL (https://huggingface.co/datasets/chayuto/klondike-llm-decisions) and link to the source repo (https://github.com/chayuto/solitaire-analytics) so readers can find the analysis context. The corpus continues to grow; pin a specific revision (load_dataset(..., revision=...)) if your work depends on a fixed snapshot.

Talk to us

Issues, comparisons, replay videos, alternative analyses are all welcome. Open an issue on the source repo or comment on the dataset discussion tab.

License

Released under CC-BY-4.0. Free to use, share, and adapt with attribution.

Card and data generated by scripts/ingest_exports.py.