RSIGame/EvoGame-SFT

EvoGame SFT

This is the cleaned and anonymized supervised fine-tuning data for EvoGame, a system of agents that generate and repair 2D games. Every config is a table the dataset viewer can display, and each row carries an ms-swift style messages column. Godot and web (Phaser + TypeScript) data are always in separate configs.

config	stage	engine	row unit	rows
plan	S2: brief → file-level implementation plan	Godot 1,979 / web 129	one plan	2,108
s1_generation_godot	S1: game-generation agent	Godot	one agent decision	66,312
s1_generation_web	S1: game-generation agent	web	one agent decision	28,097
s1_trajectories_godot / _web	S1: full anonymized trajectories (to re-window)	Godot / web	one trajectory	1,929 / 284
s3_repair_godot_positive / _negative	S3: repair prompt → patch	Godot	one repair round	1,869 / 1,861
s3_repair_agent_godot_positive / _negative	S3: multi-turn repair agent	Godot	one agent decision	48,341 / 62,961
s3_repair_web_positive / _negative	S3: repair prompt → patch	web	one branch attempt	144 / 129
s3_repair_agent_web_positive / _negative	S3: multi-turn repair agent	web	one agent decision	5,098 / 4,832

positive means a separate verifier judged the repair goal achieved. negative covers everything else (not achieved, unclear, or no verdict). Train SFT on the positive configs; keep the negative ones for preference or contrastive training.

Detailed cards for each group are in docs/: S1.md, S3_repair_godot.md, S3_repair_agent_godot.md and S3_repair_web.md.

Rules that apply to every config

• No evaluation contamination. None of the 140 benchmark tasks appear, and nothing from a baseline or evaluation run is included.
• in_s1_holdout marks tasks in the stage-1 held-out brief list. You may train on them; drop them only when evaluating on that holdout.
• Anonymized. Machine-specific paths are rewritten to /workspace/game, /workspace/harness, <path> or <assets>. User names, host names, internal service URLs, e-mail addresses, IPs and credential-shaped strings are removed. Tool outputs that listed the host's processes are never shown: the trajectory is cut before them. Every field of every row, and every document in this repo, was scanned with zero hits.
• Multi-turn rows (s1_generation_*, s3_repair_agent_*) follow one layout:
  • Roles are system, user, assistant, tool_call and tool_response, plus a tools column.
  • Each row keeps a pinned prefix (system prompt + task, and for S1 the plan when one exists), then the most recent whole decisions that fit, then the target decision last.
  • Rows are at most 32k tokens (Qwen3 chat template). Train with --loss_scale last_round.

Known recording gaps (flagged, not hidden)

Some agent runs were recorded through an SDK that stored a tool's one-line UI summary instead of the content the model actually received, e.g. Read lines 1-610 of 713 from scripts/main.gd or Listed 12 item(s).. The model saw the full content; the data only has the summary. These rows are kept and flagged:

column	meaning
n_summary_only_results	number of summary-only tool responses in the row's history
row_has_summary_only	the row's history contains at least one
target_follows_summary_only	the target decision comes right after one (the most harmful case)
trajectory_has_summary_only / session_has_summary_only	the whole trajectory or session contains one
tool_results_source (Godot repair agent)	full_chat_log (complete) or ui_summary_trajectory (summaries)
trajectory_incomplete (Godot repair agent)	the recording stops before the agent did
n_truncated_tool_outputs, session_has_truncation (web repair agent)	a very long tool output was middle-truncated

data	affected	suggested filter for clean training
s1_generation_godot	none	–
s1_generation_web	all trajectories; 20% of rows have the target right after a summary	target_follows_summary_only == false
s3_repair_agent_godot	the repair_e2 arm (b1 is complete)	tool_results_source == 'full_chat_log'
s3_repair_agent_web	most sessions	target_follows_summary_only == false

plan

Each row has three messages: a fixed system prompt, the game brief (user), and a file-level plan (assistant) in the form "Files to MODIFY / Files to CREATE", one line per file.

Every plan is grounded in a finished game:

• It was distilled from that game's real file tree.
• Every file path it references exists in that game.
• Brief and plan name the same engine.

Plans that referenced non-existent files were removed.

Columns:

• task, engine, archetype
• plan_origin (prior_grounded or prior_trajectory_extract: pre-existing plans re-verified against the game; qwen_grounded: newly distilled from the finished game)
• model, in_s1_holdout, n_chars, n_refs

s3_repair_godot_* (single-turn)

Each row is one round of multi-round self-refinement on a generated Godot game:

• [system, user = defect report and goals, assistant = unified diff]
• verdict is the verifier's judgment after the round was replayed.
• Rounds with failed sessions or empty diffs are excluded.
• Columns: task, engine, arm, round, verdict, split_role, in_s1_holdout, in_repair_holdout, n_tokens, messages. n_tokens has a median of 3.9k, and 7 rows exceed 32k.
• Final version: 3,730 rows over 345 distinct games (arm counts overlap: e2 155 + b1 242, sharing 52 games). Rounds join with s3_repair_agent_godot_* on (task, arm, round) and carry the same verdict; 6 of the 3,730 rounds have no agent session, so the join is one-to-one for 3,724 of them.

Dataset statistics

All figures below were recomputed directly from the published parquet files on 2026-09-18. Token figures are the sum of each config's stored n_tokens column (exact Qwen3 chat-template count, the same rendering used to build the rows). The plan and s1_trajectories_* configs carry no n_tokens column; their token figures were measured separately with the Qwen3 tokenizer over the concatenated messages[].content (no chat template) and are marked *.

Overview

config	stage	engine	row unit	rows	unique tasks	trajectories / sessions / rounds	tokens	median tokens	in_s1_holdout rows
plan	S2 plan	Godot + web	one plan	2,108	2,108	–	4,505,012 *	2,114.5 *	125 (5.9%)
s1_generation_godot	S1 generation	Godot	one decision	66,312	1,929	1,929 trajectories	1,702,121,541	29,565	3,554 (5.4%)
s1_generation_web	S1 generation	web	one decision	28,097	284	284 trajectories	759,953,367	30,302	5,749 (20.5%)
s1_trajectories_godot	S1 generation	Godot	one trajectory	1,929	1,929	1,929	79,551,954 *	41,341 *	95 (4.9%)
s1_trajectories_web	S1 generation	web	one trajectory	284	284	284	27,164,451 *	90,561 *	58 (20.4%)
s3_repair_godot_positive	S3 repair	Godot	one repair round	1,869	295	1,869 rounds	8,806,635	4,324	157 (8.4%)
s3_repair_godot_negative	S3 repair	Godot	one repair round	1,861	316	1,861 rounds	7,657,618	3,453	188 (10.1%)
s3_repair_agent_godot_positive	S3 repair agent	Godot	one decision	48,341	294	1,868 sessions	759,783,601	14,200	3,867 (8.0%)
s3_repair_agent_godot_negative	S3 repair agent	Godot	one decision	62,961	316	1,856 sessions	1,273,736,966	20,886	6,159 (9.8%)
s3_repair_web_positive	S3 repair	web	one branch attempt	144	42	144 attempts	1,267,624	6,005	26 (18.1%)
s3_repair_web_negative	S3 repair	web	one branch attempt	129	41	129 attempts	879,652	6,050	17 (13.2%)
s3_repair_agent_web_positive	S3 repair agent	web	one decision	5,098	42	211 sessions	129,158,687	26,534	635 (12.5%)
s3_repair_agent_web_negative	S3 repair agent	web	one decision	4,832	42	181 sessions	126,071,778	27,456	570 (11.8%)
total				223,965	2,511 (union)		4,769,437,469 (+ 111,221,417 *)		21,200

Notes on the totals:

• s1_trajectories_* (2,213 rows) are the same recordings as s1_generation_* in un-windowed form. Counting only the training-row configs gives 221,752 rows.
• Because a decision row repeats the most recent history, the 4.77 B figure counts overlapping context many times. The non-overlapping text volume of S1 is the trajectory total, 106.7 M tokens (Godot 79.6 M + web 27.2 M).
• 2,511 distinct task ids appear across the whole dataset.
• The in_s1_holdout total (21,200) is the sum of flagged rows across configs, not a count of distinct held-out tasks.
• On disk: 79 parquet and markdown files, 276,271,291 bytes (0.28 GB), plus the repository's own LFS config file.

Engine split

engine	rows	unique tasks	configs
Godot	185,252	2,189	plan (Godot part), s1_*_godot, s3_repair_godot_*, s3_repair_agent_godot_*
web (Phaser / TypeScript)	38,713	322	plan (web part), s1_*_web, s3_repair_web_*, s3_repair_agent_web_*

Godot outweighs web by 4.8:1 in rows and 6.8:1 in tasks. The two task-id sets are disjoint (0 shared ids; 2,189 + 322 = 2,511).

plan

plan_origin	rows	Godot	web
qwen_grounded (newly distilled from a finished game)	1,237	1,109	128
prior_grounded (pre-existing plan re-verified against the game)	870	870	0
prior_trajectory_extract	1	0	1
total	2,108	1,979	129

metric	all	Godot	web
rows / unique tasks	2,108 / 2,108	1,979	129
median n_chars	2,390	2,366	2,986
median n_refs (file paths referenced)	19	18	24
total n_chars	5,373,250	–	–
total n_refs	44,818	–	–
in_s1_holdout	125 (5.9%)	94 (4.8%)	31 (24.0%)
median tokens *	2,114	2,116	1,537

Every row is exactly [system, user, assistant]; 23 distinct archetypes, the largest being simulation (202), platformer (163), tower-defense (154), roguelike (137) and strategy (126).

s1_* (generation)

	Godot	web
trajectories	1,929	284
decisions (rows in s1_generation_*)	66,312	28,097
median decisions per trajectory	34	97.5
p10 / p90 decisions	20 / 52	78 / 123
has_plan trajectories	1,719 (89.1%)	112 (39.4%)
has_plan decision rows	59,299 (89.4%)	10,565 (37.6%)
in_s1_holdout trajectories	95 (4.9%)	58 (20.4%)
median n_tokens / median history tokens	29,565 / 23,101	30,302 / 20,028

Source mix (neutral source labels, trajectory level):

engine	source	trajectories	decision rows
Godot	basegame	1,039	40,177
Godot	godot_extra2	777	21,279
Godot	godot_extra1	113	4,856
web	web_base	190	19,920
web	web_expand	94	8,177

Recording-quality flags:

flag	Godot rows	web rows
row_has_summary_only	0 (0.00%)	23,865 (84.94%)
target_follows_summary_only	0 (0.00%)	5,668 (20.17%)
trajectory_has_summary_only	0 (0.00%)	28,097 (100%)

For 22 of the 284 web trajectories the stored n_decisions is one higher than the number of rows present: those trailing decisions did not fit the token budget and were dropped. Godot matches row-for-row (0 of 1,929).

s3_repair_* (single-turn, one row per repair round)

arm	engine	games	rounds	positive	negative	tokens	median tokens
repair_e2	Godot	155	1,692	1,287	405	7,589,969	–
repair_b1	Godot	242	2,038	582	1,456	8,874,284	–
Godot total		345 distinct	3,730	1,869	1,861	16,464,253	3,938.5
repair_e1_web	web	43	273 attempts	144	129	2,147,276	6,013

The two Godot arms share 52 games, so the distinct-game count is 345, not 155 + 242.

Verdict distribution:

config	verdict	rows
s3_repair_godot_positive	yes	1,869
s3_repair_godot_negative	no	1,283
	unclear	440
	none (no verdict recorded)	138
s3_repair_web_positive	fixed	144
s3_repair_web_negative	still_present	107
	not_reached	10
	does not compile	7
	unreadable	5

s3_repair_agent_* (multi-turn, one row per agent decision)

	Godot repair_b1	Godot repair_e2	Godot total	web
distinct games	242	155	345 distinct	44
sessions	2,032	1,692	3,724	392
rows	75,975	35,327	111,302	9,930
positive / negative rows	21,551 / 54,424	26,790 / 8,537	48,341 / 62,961	5,098 / 4,832
median n_tokens (pos / neg)	22,665 / 22,854	10,474 / 9,750	14,200 / 20,886	26,534 / 27,456
median n_tokens (all rows)	–	–	17,801	26,997
total tokens	1,652,458,718	381,061,849	2,033,520,567	255,230,465

Verdict distribution (Godot multi-turn): yes 48,341 · no 43,654 · unclear 15,028 · none 4,279. Web: fixed 5,098 · still_present 4,183 · not_reached 317 · does not compile 202 · unreadable 130.

Recording-quality flags, Godot multi-turn:

flag	repair_b1	repair_e2	all rows
tool_results_source == full_chat_log	75,975 (100%)	0	75,975 (68.3%)
tool_results_source == ui_summary_trajectory	0	35,327 (100%)	35,327 (31.7%)
row_has_summary_only	0	32,712 (92.6%)	32,712 (29.4%)
session_has_summary_only	0	35,287 (99.9%)	35,287 (31.7%)
trajectory_incomplete	0	29,042 (82.2%)	29,042 (26.1%)
fully clean (neither flag)	75,975 (100%)	769 (2.2%)	76,744 (69.0%)

Recording-quality flags, web multi-turn (9,930 rows):

flag	rows	share
row_has_summary_only	8,029	80.9%
target_follows_summary_only	2,995	30.2%
session_has_summary_only	9,483	95.5%
n_truncated_tool_outputs > 0	350	3.5%
in a session with any truncation	1,312	13.2%
fully clean (no summary-only, no truncation)	1,816	18.3%

Suggested clean-training filters: tool_results_source == 'full_chat_log' for Godot multi-turn, target_follows_summary_only == false for web and for S1 web.

Contamination check

The 140 distinct benchmark task ids were intersected with the task column of every config: 0 overlaps in all 13 configs, 0 across the 2,511-task union.

How this was built

The data comes from three pipeline stages run on generated 2D games: S2 plans distilled from finished game trees, S1 generation-agent recordings, and S3 repair rounds from multi-round self-refinement. Raw material was taken from the project's own run archives and agent chat logs; plans from an earlier collaborator corpus were kept only where every file path they reference exists in the matching game (870 of them survived), and the rest were re-distilled. Nothing from a baseline, ablation or evaluation run is included, and the 140 benchmark task ids are excluded from every stage — verified by an independent re-scan of the published tables. Every row, every metadata column and every document in this repository then went through a 51-check anonymization pass: absolute machine paths rewritten to /workspace/game, /workspace/harness or <path>; user, host, account and internal-service names redacted; model-gateway URLs replaced; e-mail addresses and non-loopback IPs removed; and tool outputs that dumped the build host's process table cut out of the trajectory entirely rather than scrubbed. The cleaner was checked with a planted-leak self-test and the output cross-checked with a second, independently written plain-text scanner; both report zero hits. Flawed-but-real recordings were kept rather than dropped, and carry explicit flag columns instead. Final build and verification: 2026-09-18.

Citing this dataset

@misc{evogame_sft_2026,
  title        = {EvoGame-SFT: Supervised Fine-Tuning Data for Game Generation and Repair Agents},
  author       = {Anonymous},
  year         = {2026},
  note         = {Under double-blind review; dataset released with the submission},
  howpublished = {Dataset}
}

Known limitations

• Recording gaps. One agent SDK stored a one-line UI summary of a tool result instead of the content the model actually received. This affects all 284 web S1 trajectories (84.9% of rows), the repair_e2 Godot repair arm (100% of its 35,327 rows) and most web repair sessions (95.5% of rows in a session with at least one). The rows are kept and flagged, never silently repaired.
• Truncated sessions. 26.1% of Godot multi-turn rows sit in a recording that stops before the agent did (trajectory_incomplete), all in the repair_e2 arm; 3.5% of web multi-turn rows contain a middle-truncated tool output.
• Engine imbalance. Godot dominates: 1,929 vs 284 S1 trajectories, 1,979 vs 129 plans, 345 vs 44 repaired games. Web results are far less statistically stable, and plan coverage for web trajectories is only 39%.
• Context overlap. Decision rows repeat their recent history, so row counts and summed token counts are not independent samples. For per-token accounting use the trajectory configs.
• Web history budget. Web repair rows cannot reach the intended 16k-token history: the fixed system prompt alone is ~9.9k tokens, leaving a median history of about 9.5k.
• Negative labels are heterogeneous. negative mixes "goal not achieved", "unclear" and "no verdict recorded" (138 single-turn and 4,279 multi-turn Godot rows have no verdict at all); it is not a clean failure label.
• Verifier variance. Positive/negative labels come from an automated rubric-based judge, not human review; borderline rounds can flip.
• Unreviewed model output. Every game, plan and patch is model-generated and has not been human-checked for correctness, licence or content.