Datasets:
environment stringclasses 2
values | model stringclasses 6
values | variant stringclasses 9
values | seed int64 1 3 | model_name stringclasses 6
values | model_id stringclasses 6
values | harness stringclasses 2
values | n_plays int64 0 150 | n_versions int64 18 22 | final_version stringclasses 4
values | eval_version stringclasses 3
values | eval_ps_t20_win_rate float64 0.05 0.8 โ | config_json stringlengths 222 277 | teacher_system_prompt stringclasses 9
values | actions_log stringlengths 10.2k 451k | game_ids_json stringclasses 1
value | gen_difficulty_json stringlengths 1.87k 34.1k โ | context_tokens int64 0 6.55k โ | context_file stringclasses 5
values | context_text stringclasses 5
values | raw_train stringlengths 31 62 | raw_derived stringclasses 3
values |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
puzzlescript | claude-opus-4-6 | read-write-bash-gp | 1 | Claude Opus 4.6 | claude-opus-4-6 | claude_code | 64 | 21 | v021 | v021 | 0.1875 | {
"out": "root_dir/puzzlescript/claude-opus-4-6/read-write-bash-gp_1",
"model": "claude-opus-4-6",
"tools": "gameplay",
"games": "puzzlescript",
"budget": "fixed",
"horizon": 20,
"tokens": 25000,
"context": "",
"updates": 20,
"plays": 150
}
| Your goal is to improve an agent's ability to play unfamiliar single-player games.
All games have the following attributes:
- Single player.
- At each player turn, the player takes an action, and receives feedback on the game state.
- The game has an objective and a score.
- The player plays without full knowledge of ... | [ 2.0s] orchestrator: system/init: {"cwd": "/workspace", "tools": ["Bash", "Edit", "Glob", "Grep", "Read", "Write", "mcp__play__play_own_game"], "mcp_servers": [{"name": "play", "status": "connected"}], "model": "claude-opus-4-6", "permissionMode": "bypassPermissions", "slash_commands": ["deep-research", "dataviz",... | {} | {
"gen_test_game_L0": {
"game": "test_game.ps",
"level": 0,
"algo": "bfs",
"n_rules": 4,
"n_objects": 5,
"width": 5,
"height": 6,
"iters": 25,
"budget_hit": false,
"solved": true,
"claimed_len": 6,
"actions": [
0,
1,
0,
0,
3,
2
],... | null | null | null | puzzlescript/claude-opus-4-6/read-write-bash-gp_1/train.tar.gz | null |
puzzlescript | claude-opus-4-6 | read-write-bash-gp | 2 | Claude Opus 4.6 | claude-opus-4-6 | claude_code | 86 | 21 | v021 | v021 | 0.25 | {
"out": "root_dir/puzzlescript/claude-opus-4-6/read-write-bash-gp_2",
"model": "claude-opus-4-6",
"tools": "gameplay",
"games": "puzzlescript",
"budget": "fixed",
"horizon": 20,
"tokens": 25000,
"context": "",
"updates": 20,
"plays": 150
}
| Your goal is to improve an agent's ability to play unfamiliar single-player games.
All games have the following attributes:
- Single player.
- At each player turn, the player takes an action, and receives feedback on the game state.
- The game has an objective and a score.
- The player plays without full knowledge of ... | [ 1.9s] orchestrator: system/init: {"cwd": "/workspace", "tools": ["Bash", "Edit", "Glob", "Grep", "Read", "Write", "mcp__play__play_own_game"], "mcp_servers": [{"name": "play", "status": "connected"}], "model": "claude-opus-4-6", "permissionMode": "bypassPermissions", "slash_commands": ["deep-research", "dataviz",... | {} | {
"gen_test_game_L0": {
"game": "test_game.puzzlescript",
"level": 0,
"algo": "bfs",
"n_rules": 4,
"n_objects": 5,
"width": 5,
"height": 6,
"iters": 25,
"budget_hit": false,
"solved": true,
"claimed_len": 6,
"actions": [
0,
1,
0,
0,
3,
... | null | null | null | puzzlescript/claude-opus-4-6/read-write-bash-gp_2/train.tar.gz | null |
puzzlescript | claude-opus-4-6 | read-write-bash-gp | 3 | Claude Opus 4.6 | claude-opus-4-6 | claude_code | 123 | 21 | v021 | v021 | 0.1375 | "{\n \"out\": \"root_dir/puzzlescript/claude-opus-4-6/read-write-bash-gp_3\",\n \"model\": \"claud(...TRUNCATED) | "Your goal is to improve an agent's ability to play unfamiliar single-player games. \nAll games have(...TRUNCATED) | "[ 2.0s] orchestrator: system/init: {\"cwd\": \"/workspace\", \"tools\": [\"Bash\", \"Edit\", \"(...TRUNCATED) | {} | "{\n \"gen_test_game_L0\": {\n \"game\": \"test_game.puzzlescript\",\n \"level\": 0,\n \"a(...TRUNCATED) | null | null | null | puzzlescript/claude-opus-4-6/read-write-bash-gp_3/train.tar.gz | null |
puzzlescript | claude-opus-5 | read-write-bash-gp | 1 | Claude Opus 5 | claude-opus-5 | claude_code | 75 | 21 | v021 | v021 | 0.5 | "{\n \"out\": \"root_dir/puzzlescript/claude-opus-5/read-write-bash-gp_1\",\n \"model\": \"claude-(...TRUNCATED) | "Your goal is to improve an agent's ability to play unfamiliar single-player games. \nAll games have(...TRUNCATED) | "[ 2.0s] orchestrator: system/init: {\"cwd\": \"/workspace\", \"tools\": [\"Bash\", \"Edit\", \"(...TRUNCATED) | {} | "{\n \"gen_sokoban_L0\": {\n \"game\": \"sokoban.txt\",\n \"level\": 0,\n \"algo\": \"bfs\(...TRUNCATED) | null | null | null | puzzlescript/claude-opus-5/read-write-bash-gp_1/train.tar.gz | puzzlescript/claude-opus-5/read-write-bash-gp_1/derived_game_sources.tar.gz |
puzzlescript | claude-opus-5 | read-write-bash-gp | 2 | Claude Opus 5 | claude-opus-5 | claude_code | 82 | 21 | v021 | v021 | 0.4125 | "{\n \"out\": \"root_dir/puzzlescript/claude-opus-5/read-write-bash-gp_2\",\n \"model\": \"claude-(...TRUNCATED) | "Your goal is to improve an agent's ability to play unfamiliar single-player games. \nAll games have(...TRUNCATED) | "[ 1.9s] orchestrator: system/init: {\"cwd\": \"/workspace\", \"tools\": [\"Bash\", \"Edit\", \"(...TRUNCATED) | {} | "{\n \"gen_g1_L0\": {\n \"game\": \"g1.txt\",\n \"level\": 0,\n \"algo\": \"bfs\",\n \"(...TRUNCATED) | null | null | null | puzzlescript/claude-opus-5/read-write-bash-gp_2/train.tar.gz | puzzlescript/claude-opus-5/read-write-bash-gp_2/derived_game_sources.tar.gz |
puzzlescript | claude-opus-5 | read-write-bash-gp | 3 | Claude Opus 5 | claude-opus-5 | claude_code | 82 | 21 | v021 | v021 | 0.4375 | "{\n \"out\": \"root_dir/puzzlescript/claude-opus-5/read-write-bash-gp_3\",\n \"model\": \"claude-(...TRUNCATED) | "Your goal is to improve an agent's ability to play unfamiliar single-player games. \nAll games have(...TRUNCATED) | "[ 2.0s] orchestrator: system/init: {\"cwd\": \"/workspace\", \"tools\": [\"Bash\", \"Edit\", \"(...TRUNCATED) | {} | "{\n \"gen_sokoban_L0\": {\n \"game\": \"sokoban.txt\",\n \"level\": 0,\n \"algo\": \"bfs\(...TRUNCATED) | null | null | null | puzzlescript/claude-opus-5/read-write-bash-gp_3/train.tar.gz | puzzlescript/claude-opus-5/read-write-bash-gp_3/derived_game_sources.tar.gz |
puzzlescript | claude-opus-5 | read-write-bash | 1 | Claude Opus 5 | claude-opus-5 | claude_code | 0 | 22 | v021 | v021 | 0.4125 | "{\n \"out\": \"root_dir/puzzlescript/claude-opus-5/read-write-bash_1\",\n \"model\": \"claude-opu(...TRUNCATED) | "Your goal is to improve an agent's ability to play unfamiliar single-player games. \nAll games have(...TRUNCATED) | "[ 1.9s] orchestrator: system/init: {\"cwd\": \"/workspace\", \"tools\": [\"Bash\", \"Edit\", \"(...TRUNCATED) | {} | null | null | null | null | puzzlescript/claude-opus-5/read-write-bash_1/train.tar.gz | null |
puzzlescript | claude-opus-5 | read-write-bash | 2 | Claude Opus 5 | claude-opus-5 | claude_code | 0 | 22 | v021 | v021 | 0.35 | "{\n \"out\": \"root_dir/puzzlescript/claude-opus-5/read-write-bash_2\",\n \"model\": \"claude-opu(...TRUNCATED) | "Your goal is to improve an agent's ability to play unfamiliar single-player games. \nAll games have(...TRUNCATED) | "[ 1.9s] orchestrator: system/init: {\"cwd\": \"/workspace\", \"tools\": [\"Bash\", \"Edit\", \"(...TRUNCATED) | {} | null | null | null | null | puzzlescript/claude-opus-5/read-write-bash_2/train.tar.gz | null |
puzzlescript | claude-opus-5 | read-write-bash | 3 | Claude Opus 5 | claude-opus-5 | claude_code | 0 | 22 | v021 | v021 | 0.3125 | "{\n \"out\": \"root_dir/puzzlescript/claude-opus-5/read-write-bash_3\",\n \"model\": \"claude-opu(...TRUNCATED) | "Your goal is to improve an agent's ability to play unfamiliar single-player games. \nAll games have(...TRUNCATED) | "[ 1.9s] orchestrator: system/init: {\"cwd\": \"/workspace\", \"tools\": [\"Bash\", \"Edit\", \"(...TRUNCATED) | {} | null | null | null | null | puzzlescript/claude-opus-5/read-write-bash_3/train.tar.gz | null |
puzzlescript | claude-opus-5 | read-write | 1 | Claude Opus 5 | claude-opus-5 | claude_code | 0 | 21 | v021 | v021 | 0.3125 | "{\n \"out\": \"root_dir/puzzlescript/claude-opus-5/read-write_1\",\n \"model\": \"claude-opus-5\"(...TRUNCATED) | "Your goal is to improve an agent's ability to play unfamiliar single-player games. \nAll games have(...TRUNCATED) | "[ 1.9s] orchestrator: system/init: {\"cwd\": \"/workspace\", \"tools\": [\"Edit\", \"Glob\", \"(...TRUNCATED) | {} | null | null | null | null | puzzlescript/claude-opus-5/read-write_1/train.tar.gz | null |
Teacher runs and evaluations
Datasets and experiment logs from Do LLM Agents Still Need to Be Told How to Self-Improve?
This dataset has the 42 teacher runs from the paper's main results, held-out evaluations for each teacher run, baseline runs and benchmark datasets.
Layout
benchmarks/ Evaluation benchmarks (ps_t20, ps_t30, ps_narrow, arcagi3, mazebench)
puzzlescript/<model>/
baselines/ Baseline evaluations (player with the default prompt)
read-write_<seed>/ Experiments: Teacher has Read/Write
read-write-bash_<seed>/ Experiments: Teacher has Read/Write + Bash
read-write-bash-gp_<seed>/ Experiments: Teacher has Read/Write + Bash + Gameplay (PuzzleScript)
python/
context-<tokens>_<seed>/ Experiments: Teacher has Read/Write + Bash + Gameplay (Python) with varying amounts of extra context about the evaluation setting.
game_context/ Evaluation setting descriptions used in the context-<tokens>_<seed> experiments
Every experiment has three seeds. read-write-bash-gp is the paper's canonical setting and exists for all six
models (claude-opus-5, claude-opus-4-6, gpt-6-astra, gpt-5.6-sol, glm-5.3, glm-4.7); the
other settings are claude-opus-5 only.
A run folder
train.tar.gz Logs from the teacher session
evals/<benchmark>.tar.gz Evaluations of optimized context on held-out benchmarks
tables/*.parquet Logs from the teacher session formatted as tables
The .tar.gz archives hold the raw experiment files. The
parquet files in tables/ arrange the contents as rows (one row per
evaluation, per game, per context file, ...), so they can be loaded easily.
A teacher run contains:
config.json Experiment settings
actions.log Timeline of the teacher's tool calls, gameplays, and context updates
library_versions/ Every intermediate version of the optimized text artifacts (v001 is the player default prompt)
player_folder/ Final version of text artifacts
plays.jsonl, plays/ Teacher-generated games and observed gameplays
sandbox/ The teacher's working directory
teacher_transcript.json The teacher's full session in order: its prompts, reasoning, tool calls and tool results
Tables take the following form:
| Table | One row per |
|---|---|
runs |
teacher run: settings, actions.log, final ps_t20 win rate |
evals |
evaluation: benchmark, context version, win rate |
eval_plays |
evaluation game, with the full transcript |
train_plays |
game the teacher had the player play, with transcript and game source |
context_versions |
file in one version of the player's context |
sandbox_files |
file in the teacher's working directory |
teacher_transcript |
message in the teacher's session: prompt, reasoning, tool call or tool result |
An evaluation folder contains evals/<benchmark>/<version>/ with config.json, report.json (the win
rate), plays.jsonl and plays/ (every game with its full transcript).
Baselines
Baseline runs (results on the evaluation benchmarks with the default player prompt) are in puzzlescript/<model>/baselines/<benchmark>_<seed>.tar.gz.
- Downloads last month
- -