Datasets:

efederici
/

smol-software-harbor

id string	codebase string	language string	domain string	task_type string	instruction_language string	difficulty int64	split string	files_touched list	oracle_steps int64	n_attacks int64	n_files int64	toolchains list	path string	search bool
c-arena-alloc--align-allocations	c-arena-alloc	c	memory arena allocator	fix_bug	en	2	new_train	[ "src/arena.c" ]	1	1	5	[ "cc", "make" ]	tasks/c-arena-alloc--align-allocations	false
c-arg-tokenizer--backslash-escape	c-arg-tokenizer	c	shell-like argument splitting	fix_bug	en	2	new_train	[ "src/tok.c" ]	1	1	5	[ "cc", "make" ]	tasks/c-arg-tokenizer--backslash-escape	false
c-arg-tokenizer--test-quotes	c-arg-tokenizer	c	shell-like argument splitting	add_test	en	2	new_train	[ "tests/test_main.c" ]	1	1	5	[ "cc", "make" ]	tasks/c-arg-tokenizer--test-quotes	false
c-audio-plugins--answer-max-chain	c-audio-plugins	c	audio filter plugins	search_answer	en	1	new_train	[ "answer.txt" ]	1	2	13	[ "cc", "make" ]	tasks/c-audio-plugins--answer-max-chain	true
c-audio-plugins--negative-clipping	c-audio-plugins	c	audio filter plugins	locate_fix	de	2	new_train	[ "src/plugins/gain.c" ]	1	2	13	[ "cc", "make" ]	tasks/c-audio-plugins--negative-clipping	true
c-base64--encode-padding	c-base64	c	base64 encoding	fix_bug	en	1	new_train	[ "src/b64.c" ]	1	1	5	[ "cc", "make" ]	tasks/c-base64--encode-padding	false
c-base64--implement-decode	c-base64	c	base64 encoding	implement_stub	en	2	new_train	[ "src/b64.c" ]	1	2	5	[ "cc", "make" ]	tasks/c-base64--implement-decode	false
c-bitset--implement-bitset	c-bitset	c	bit manipulation	implement_stub	de	2	train	[ "src/bitset.c" ]	1	2	5	[ "cc", "make" ]	tasks/c-bitset--implement-bitset	null
c-bloom-filter--probe-range	c-bloom-filter	c	Bloom filter	fix_bug	en	1	new_train	[ "src/bloom.c" ]	1	1	5	[ "cc", "make" ]	tasks/c-bloom-filter--probe-range	false
c-config-kv--remove-legacy-lookup	c-config-kv	c	configuration files (C)	remove_dead_code	de	1	test_ood	[ "src/config.c" ]	1	2	5	[ "cc", "make" ]	tasks/c-config-kv--remove-legacy-lookup	null
c-config-kv--windows-line-ends	c-config-kv	c	configuration files (C)	fix_bug	en	1	test_ood	[ "src/config.c" ]	1	1	5	[ "cc", "make" ]	tasks/c-config-kv--windows-line-ends	null
c-conv-cli--precision-flag	c-conv-cli	c	unit converter CLI	add_cli_flag	es	2	new_train	[ "conv.c" ]	1	2	2	[ "cc", "make" ]	tasks/c-conv-cli--precision-flag	false
c-crc32--implement-crc32	c-crc32	c	file checksums (CRC-32)	implement_stub	en	2	new_train	[ "src/crc32.c" ]	1	2	5	[ "cc", "make" ]	tasks/c-crc32--implement-crc32	false
c-csv-colsum--cannot-open	c-csv-colsum	c	CSV processing	handle_missing_file	en	1	train	[ "src/main.c" ]	1	2	5	[ "cc", "make" ]	tasks/c-csv-colsum--cannot-open	null
c-csv-colsum--fix-makefile-link	c-csv-colsum	c	CSV processing	fix_build	es	1	train	[ "Makefile" ]	1	2	5	[ "cc", "make" ]	tasks/c-csv-colsum--fix-makefile-link	null
c-date-utils--february-in-leap-years	c-date-utils	c	calendar dates	edge_case	en	1	train	[ "src/dates.c" ]	1	2	6	[ "cc", "make" ]	tasks/c-date-utils--february-in-leap-years	null
c-date-utils--rename-is-leap	c-date-utils	c	calendar dates	rename_symbol	en	2	train	[ "src/dates.c", "src/dates.h" ]	4	2	6	[ "cc", "make" ]	tasks/c-date-utils--rename-is-leap	null
c-edit-distance--first-column	c-edit-distance	c	spelling suggestions (edit distance)	fix_bug	fr	1	new_train	[ "src/lev.c" ]	1	1	8	[ "cc", "make" ]	tasks/c-edit-distance--first-column	false
c-edit-distance--test-suggest	c-edit-distance	c	spelling suggestions (edit distance)	add_test	en	2	new_train	[ "tests/test_main.c" ]	1	1	8	[ "cc", "make" ]	tasks/c-edit-distance--test-suggest	false
c-hash-table--answer-longest-chain	c-hash-table	c	hash tables	read_answer	en	2	train	[ "answer.txt" ]	1	2	6	[ "cc", "make" ]	tasks/c-hash-table--answer-longest-chain	null
c-hash-table--chain-lookup	c-hash-table	c	hash tables	fix_bug	it	2	train	[ "src/table.c" ]	1	2	6	[ "cc", "make" ]	tasks/c-hash-table--chain-lookup	null
c-histogram--empty-input	c-histogram	c	terminal bar charts	edge_case	en	1	new_train	[ "histo.c" ]	1	1	2	[ "cc", "make" ]	tasks/c-histogram--empty-input	false
c-histogram--width-flag	c-histogram	c	terminal bar charts	add_cli_flag	de	2	new_train	[ "histo.c" ]	1	2	2	[ "cc", "make" ]	tasks/c-histogram--width-flag	false
c-ipv4-parse--format-order	c-ipv4-parse	c	IPv4 address parsing	fix_bug	en	1	new_train	[ "src/ipv4.c" ]	1	1	5	[ "cc", "make" ]	tasks/c-ipv4-parse--format-order	false
c-ipv4-parse--strict-parse	c-ipv4-parse	c	IPv4 address parsing	add_validation	fr	2	new_train	[ "src/ipv4.c" ]	1	2	5	[ "cc", "make" ]	tasks/c-ipv4-parse--strict-parse	false
c-iso-week--from-iso-week	c-iso-week	c	ISO-8601 week numbers	add_feature	es	2	new_train	[ "src/isoweek.c", "src/isoweek.h" ]	2	2	5	[ "cc", "make" ]	tasks/c-iso-week--from-iso-week	false
c-iso-week--last-days-of-year	c-iso-week	c	ISO-8601 week numbers	fix_bug	en	2	new_train	[ "src/isoweek.c" ]	1	1	5	[ "cc", "make" ]	tasks/c-iso-week--last-days-of-year	false
c-job-heap--sift-smaller-child	c-job-heap	c	priority job queue	fix_bug	en	2	new_test	[ "src/heap.c" ]	1	1	5	[ "cc", "make" ]	tasks/c-job-heap--sift-smaller-child	false
c-led-scroller--digit-glyphs	c-led-scroller	c	LED matrix clock firmware	locate_fix	it	2	new_train	[ "src/font.c" ]	1	2	11	[ "cc", "make" ]	tasks/c-led-scroller--digit-glyphs	true
c-led-scroller--slower-scroll	c-led-scroller	c	LED matrix clock firmware	find_and_change	en	1	new_train	[ "src/config.h" ]	1	2	11	[ "cc", "make" ]	tasks/c-led-scroller--slower-scroll	true
c-luhn--implement-luhn	c-luhn	c	card number checksums	implement_stub	en	2	train	[ "src/luhn.c" ]	1	2	5	[ "cc", "make" ]	tasks/c-luhn--implement-luhn	null
c-luhn--test-luhn-edges	c-luhn	c	card number checksums	add_test	fr	2	train	[ "tests/test_main.c" ]	1	1	5	[ "cc", "make" ]	tasks/c-luhn--test-luhn-edges	null
c-matrix--fix-det-sign	c-matrix	c	linear algebra	fix_bug	fr	2	train	[ "src/matrix.c" ]	1	2	5	[ "cc", "make" ]	tasks/c-matrix--fix-det-sign	null
c-matrix--fix-missing-header	c-matrix	c	linear algebra	fix_build	en	1	train	[ "src/matrix.c" ]	1	1	5	[ "cc", "make" ]	tasks/c-matrix--fix-missing-header	null
c-moving-average--implement-push	c-moving-average	c	sensor smoothing (moving average)	implement_stub	it	2	new_train	[ "src/avg.c" ]	1	2	5	[ "cc", "make" ]	tasks/c-moving-average--implement-push	false
c-path-normalize--leading-dotdot	c-path-normalize	c	file path normalisation	edge_case	en	1	new_train	[ "src/path.c" ]	1	1	5	[ "cc", "make" ]	tasks/c-path-normalize--leading-dotdot	false
c-path-normalize--test-normalize	c-path-normalize	c	file path normalisation	add_test	en	2	new_train	[ "tests/test_main.c" ]	1	1	5	[ "cc", "make" ]	tasks/c-path-normalize--test-normalize	false
c-queue-sim--answer-monday-wait	c-queue-sim	c	bank queue simulation	read_answer	fr	1	new_train	[ "answer.txt" ]	1	2	6	[ "cc", "make" ]	tasks/c-queue-sim--answer-monday-wait	false
c-ring-buffer--fix-wraparound	c-ring-buffer	c	queues and ring buffers	fix_bug	en	2	train	[ "src/ring.c" ]	1	2	5	[ "cc", "make" ]	tasks/c-ring-buffer--fix-wraparound	null
c-ring-buffer--test-overwrite	c-ring-buffer	c	queues and ring buffers	add_test	en	2	train	[ "tests/test_main.c" ]	1	1	5	[ "cc", "make" ]	tasks/c-ring-buffer--test-overwrite	null
c-rpn-calc--division-by-zero	c-rpn-calc	c	calculators	add_validation	en	1	test_seen_types	[ "src/rpn.c", "tests/test_main.c" ]	2	2	5	[ "cc", "make" ]	tasks/c-rpn-calc--division-by-zero	null
c-rpn-calc--implement-rpn-eval	c-rpn-calc	c	calculators	implement_stub	en	3	test_seen_types	[ "src/rpn.c" ]	1	2	5	[ "cc", "make" ]	tasks/c-rpn-calc--implement-rpn-eval	null
c-sensor-firmware--answer-bus-speed-file	c-sensor-firmware	c	sensor firmware	search_answer	es	1	new_train	[ "answer.txt" ]	1	3	15	[ "cc", "make" ]	tasks/c-sensor-firmware--answer-bus-speed-file	true
c-sensor-firmware--crc-polynomial	c-sensor-firmware	c	sensor firmware	locate_fix	en	2	new_train	[ "src/util/crc8.c" ]	1	2	15	[ "cc", "make" ]	tasks/c-sensor-firmware--crc-polynomial	true
c-seven-segment--clean-and-check-targets	c-seven-segment	c	seven-segment display	add_make_target	en	2	train	[ "Makefile", "tests/42.txt" ]	3	2	3	[ "cc", "make" ]	tasks/c-seven-segment--clean-and-check-targets	null
c-stock-ticker--missing-type	c-stock-ticker	c	stock price ticker	fix_build	en	1	new_train	[ "src/fmt.h" ]	1	2	7	[ "cc", "make" ]	tasks/c-stock-ticker--missing-type	false
c-strbuf--room-for-nul	c-strbuf	c	growable string buffer	edge_case	en	2	new_train	[ "src/strbuf.c" ]	1	1	5	[ "cc", "make" ]	tasks/c-strbuf--room-for-nul	false
c-strutils--fix-reverse-off-by-one	c-strutils	c	string utilities	fix_bug	en	1	train	[ "src/strutils.c" ]	1	2	5	[ "cc", "make" ]	tasks/c-strutils--fix-reverse-off-by-one	null
c-strutils--trim-all-spaces	c-strutils	c	string utilities	edge_case	de	2	train	[ "src/strutils.c" ]	1	2	5	[ "cc", "make" ]	tasks/c-strutils--trim-all-spaces	null
c-tempconv--clean-target	c-tempconv	c	unit conversion (temperature)	add_make_target	it	1	train	[ "Makefile" ]	2	3	5	[ "cc", "make" ]	tasks/c-tempconv--clean-target	null
c-tempconv--fix-link	c-tempconv	c	unit conversion (temperature)	fix_build	en	1	train	[ "Makefile" ]	1	2	5	[ "cc", "make" ]	tasks/c-tempconv--fix-link	null
c-token-bucket--init-validation	c-token-bucket	c	API rate limiting	add_validation	de	1	new_train	[ "src/bucket.c", "src/bucket.h" ]	3	2	5	[ "cc", "make" ]	tasks/c-token-bucket--init-validation	false
c-token-bucket--take-tokens	c-token-bucket	c	API rate limiting	implement_stub	en	2	new_train	[ "src/bucket.c" ]	1	3	5	[ "cc", "make" ]	tasks/c-token-bucket--take-tokens	false
c-utf8-length--count-code-points	c-utf8-length	c	UTF-8 string length	fix_bug	en	2	new_test	[ "src/u8.c" ]	1	2	5	[ "cc", "make" ]	tasks/c-utf8-length--count-code-points	false
c-wordcount--count-word-runs	c-wordcount	c	text statistics (stdin)	edge_case	en	1	train	[ "wc.c" ]	1	1	2	[ "cc", "make" ]	tasks/c-wordcount--count-word-runs	null
c-wordcount--lines-flag	c-wordcount	c	text statistics (stdin)	add_cli_flag	es	2	train	[ "wc.c" ]	2	1	2	[ "cc", "make" ]	tasks/c-wordcount--lines-flag	null
go-bowling--strike-bonus	go-bowling	go	sports scoring	fix_bug	en	2	train	[ "bowling.go" ]	1	2	3	[ "go" ]	tasks/go-bowling--strike-bonus	null
go-expense-ledger--cannot-open	go-expense-ledger	go	expense tracking	handle_missing_file	en	1	test_ood	[ "cmd/ledger/main.go" ]	1	2	5	[ "go" ]	tasks/go-expense-ledger--cannot-open	null
go-expense-ledger--sorted-report	go-expense-ledger	go	expense tracking	make_deterministic	en	1	test_ood	[ "ledger/ledger.go" ]	2	1	5	[ "go" ]	tasks/go-expense-ledger--sorted-report	null
go-greet-cli--times-validation	go-greet-cli	go	command-line flags	fix_bug	de	1	train	[ "main.go" ]	1	2	3	[ "go" ]	tasks/go-greet-cli--times-validation	null
go-greet-cli--upper-flag	go-greet-cli	go	command-line flags	add_cli_flag	en	2	train	[ "main.go" ]	2	2	3	[ "go" ]	tasks/go-greet-cli--upper-flag	null
go-maze-bfs--implement-shortest-path	go-maze-bfs	go	maze solving	implement_stub	en	3	train	[ "maze/maze.go" ]	1	2	3	[ "go" ]	tasks/go-maze-bfs--implement-shortest-path	null
go-notifier-config--timeout-setting	go-notifier-config	go	service configuration (JSON)	config_change	en	2	train	[ "config.json", "config/config.go" ]	3	2	4	[ "go" ]	tasks/go-notifier-config--timeout-setting	null
go-notifier-config--validate-port	go-notifier-config	go	service configuration (JSON)	add_validation	it	2	train	[ "config/config.go" ]	1	2	4	[ "go" ]	tasks/go-notifier-config--validate-port	null
go-postcodes--fix-postcode-regex	go-postcodes	go	postal code validation	fix_regex	fr	2	test_ood	[ "postcode/postcode.go" ]	1	2	3	[ "go" ]	tasks/go-postcodes--fix-postcode-regex	null
go-reorder-points--fix-type-mismatch	go-reorder-points	go	stock replenishment	fix_build	en	1	test_seen_types	[ "reorder/reorder.go" ]	1	1	3	[ "go" ]	tasks/go-reorder-points--fix-type-mismatch	null
go-reorder-points--use-reorder-point	go-reorder-points	go	stock replenishment	refactor_extract	en	2	test_seen_types	[ "reorder/reorder.go" ]	2	2	3	[ "go" ]	tasks/go-reorder-points--use-reorder-point	null
go-round-robin--answer-leader	go-round-robin	go	tournament standings	read_answer	fr	2	train	[ "answer.txt" ]	1	2	4	[ "go" ]	tasks/go-round-robin--answer-leader	null
go-round-robin--away-win-points	go-round-robin	go	tournament standings	fix_bug	en	1	train	[ "standings/standings.go" ]	1	1	4	[ "go" ]	tasks/go-round-robin--away-win-points	null
go-runes--reverse-runes	go-runes	go	unicode text	edge_case	en	1	test_seen_types	[ "text/text.go" ]	1	2	3	[ "go" ]	tasks/go-runes--reverse-runes	null
go-runes--test-truncate	go-runes	go	unicode text	add_test	it	2	test_seen_types	[ "text/text_test.go" ]	1	1	3	[ "go" ]	tasks/go-runes--test-truncate	null
go-semver--release-1-5-0	go-semver	go	semantic version parsing	bump_version	es	3	train	[ "CHANGELOG.md", "VERSION", "semver/semver.go" ]	3	2	6	[ "go" ]	tasks/go-semver--release-1-5-0	null
go-semver--rename-cmp-compare	go-semver	go	semantic version parsing	rename_symbol	en	2	train	[ "semver/latest.go", "semver/semver.go", "semver/semver_test.go" ]	3	2	6	[ "go" ]	tasks/go-semver--rename-cmp-compare	null
go-shortlink--implement-decode	go-shortlink	go	URL shortener	implement_stub	en	2	train	[ "base62/base62.go" ]	1	2	3	[ "go" ]	tasks/go-shortlink--implement-decode	null
go-shortlink--test-encode-boundaries	go-shortlink	go	URL shortener	add_test	en	2	train	[ "base62/base62_test.go" ]	1	1	3	[ "go" ]	tasks/go-shortlink--test-encode-boundaries	null
go-ttl-store--expiry-instant	go-ttl-store	go	caching with expiry	fix_bug	en	1	train	[ "store/store.go" ]	1	2	3	[ "go" ]	tasks/go-ttl-store--expiry-instant	null
go-ttl-store--live-keys	go-ttl-store	go	caching with expiry	add_feature	es	2	train	[ "store/store.go" ]	2	2	3	[ "go" ]	tasks/go-ttl-store--live-keys	null
go-wordwrap--exact-width-and-long-words	go-wordwrap	go	text layout	edge_case	en	2	train	[ "wrap/wrap.go" ]	1	2	4	[ "go" ]	tasks/go-wordwrap--exact-width-and-long-words	null
go-wordwrap--fix-unused-import	go-wordwrap	go	text layout	fix_build	de	1	train	[ "cmd/wrap/main.go" ]	1	1	4	[ "go" ]	tasks/go-wordwrap--fix-unused-import	null
java-inventory-csv--answer-most-valuable	java-inventory-csv	java	inventory valuation	read_answer	fr	1	train	[ "answer.txt" ]	1	2	5	[ "java", "javac" ]	tasks/java-inventory-csv--answer-most-valuable	null
java-inventory-csv--missing-import	java-inventory-csv	java	inventory valuation	fix_build	en	1	train	[ "src/Inventory.java" ]	1	1	5	[ "java", "javac" ]	tasks/java-inventory-csv--missing-import	null
java-parking-lot--free-the-spot	java-parking-lot	java	parking lot	fix_bug	en	1	train	[ "src/ParkingLot.java" ]	1	2	3	[ "java", "javac" ]	tasks/java-parking-lot--free-the-spot	null
java-roman-numerals--implement-to-roman	java-roman-numerals	java	roman numerals	implement_stub	en	2	train	[ "src/Roman.java" ]	1	2	3	[ "java", "javac" ]	tasks/java-roman-numerals--implement-to-roman	null
java-roman-numerals--test-from-roman	java-roman-numerals	java	roman numerals	add_test	it	2	train	[ "test/TestMain.java" ]	1	1	3	[ "java", "javac" ]	tasks/java-roman-numerals--test-from-roman	null
java-savings-account--compound-interest	java-savings-account	java	savings accounts	fix_bug	de	2	train	[ "src/Account.java" ]	1	2	3	[ "java", "javac" ]	tasks/java-savings-account--compound-interest	null
java-savings-account--positive-deposits	java-savings-account	java	savings accounts	add_validation	en	1	train	[ "src/Account.java" ]	1	2	3	[ "java", "javac" ]	tasks/java-savings-account--positive-deposits	null
java-unit-converter--decimals-and-miles	java-unit-converter	java	unit conversion (Java)	config_change	en	3	test_seen_types	[ "src/Converter.java", "test/TestMain.java", "units.properties" ]	5	2	3	[ "java", "javac" ]	tasks/java-unit-converter--decimals-and-miles	null
java-word-frequency--alphabetical-ties	java-word-frequency	java	word frequency	make_deterministic	es	1	test_ood	[ "src/WordFreq.java" ]	1	2	3	[ "java", "javac" ]	tasks/java-word-frequency--alphabetical-ties	null
js-cart-discounts--bogo-coupon	js-cart-discounts	javascript	shopping cart discounts	add_feature	es	2	train	[ "src/cart.js" ]	1	2	3	[ "node" ]	tasks/js-cart-discounts--bogo-coupon	null
js-cart-discounts--percent-before-fixed	js-cart-discounts	javascript	shopping cart discounts	fix_bug	en	2	train	[ "src/cart.js" ]	1	2	3	[ "node" ]	tasks/js-cart-discounts--percent-before-fixed	null
js-cart-totals--code-normalise	js-cart-totals	javascript	web shop checkout totals	edge_case	en	1	new_train	[ "src/discounts.js" ]	1	2	6	[ "node" ]	tasks/js-cart-totals--code-normalise	false
js-cart-totals--free-shipping	js-cart-totals	javascript	web shop checkout totals	add_feature	en	3	new_train	[ "src/discounts.js", "src/shipping.js", "src/totals.js" ]	6	3	6	[ "node" ]	tasks/js-cart-totals--free-shipping	false
js-chat-commands--answer-max-mute	js-chat-commands	javascript	chat server slash commands	search_answer	en	1	new_train	[ "answer.txt" ]	1	2	7	[ "node" ]	tasks/js-chat-commands--answer-max-mute	true
js-chat-commands--mute-permission	js-chat-commands	javascript	chat server slash commands	locate_fix	en	1	new_train	[ "src/commands/mute.js" ]	1	2	7	[ "node" ]	tasks/js-chat-commands--mute-permission	true
js-chess-clock--flag-fall	js-chess-clock	javascript	chess clocks	edge_case	en	1	new_train	[ "clock.js" ]	1	2	2	[ "node" ]	tasks/js-chess-clock--flag-fall	false
js-chess-clock--implement-press	js-chess-clock	javascript	chess clocks	implement_stub	en	2	new_train	[ "clock.js" ]	1	2	2	[ "node" ]	tasks/js-chess-clock--implement-press	false
js-circuit-breaker--half-open	js-circuit-breaker	javascript	circuit breaker pattern	fix_bug	en	2	new_train	[ "src/breaker.js" ]	1	2	2	[ "node" ]	tasks/js-circuit-breaker--half-open	false
js-color-contrast--linear-segment	js-color-contrast	javascript	colour contrast (WCAG)	fix_bug	en	2	new_train	[ "src/contrast.js" ]	1	2	2	[ "node" ]	tasks/js-color-contrast--linear-segment	false
js-color-convert--implement-rgb-to-hex	js-color-convert	javascript	colour conversion	implement_stub	en	1	train	[ "lib/color.js" ]	1	2	3	[ "node" ]	tasks/js-color-convert--implement-rgb-to-hex	null
js-color-convert--shorthand-hex	js-color-convert	javascript	colour conversion	edge_case	es	2	train	[ "lib/color.js" ]	1	2	3	[ "node" ]	tasks/js-color-convert--shorthand-hex	null

End of preview. Expand in Data Studio

🛠️ smol-software-harbor

1,002 small, verifiable software-engineering tasks on 576 hand-written mini-codebases in 12 languages, for small coding agents: read, find the right place, edit, run the tests, stop.

Ready for RL with palingenesis: pgs harbor check validates every task and pgs harbor rows converts them into training rows (see Train with palingenesis).

Most agentic coding benchmarks are built for large models: big repositories, long horizons, few tasks per skill. smol-software-harbor goes the other way. Every task sits in its own small, hand-written codebase (4 files median, up to 15), with a different structure, domain, naming style and logic each time. No codebase carries more than 2 tasks, so the set teaches habits rather than templates. A typical task means reading 1–3 files, editing 1–2 and running the project's tests, in 3–10 tool calls. 89% of the reference solutions touch a single file and 86% are a single edit.

About a fifth of the tasks are search tasks: the instruction does not say where the problem is. The codebase holds distractors (similar names, dead code, vendored copies, a second config), so blind edits fail and the agent has to locate the code first (grep, find, follow a traceback).

Reward is all-or-nothing: hidden tests + the visible project tests + collateral-damage checks. No LLM in the reward path.

At a glance


tasks	1,002
codebases	576 (563 distinct domains: warehouse reports, parking lots, plant care, audio plugins, …)
programming languages	python 534 · javascript 127 · shell 79 · c 56 · ruby 46 · sql 46 · rust 39 · go 23 · lua 17 · perl 17 · java 9 · php 9
instruction languages	en 685 · es 83 · de 82 · fr 77 · it 75 (natural per-language text, not machine-translated)
difficulty (1–3)	421 / 507 / 74
search tasks	225
attacks (plausible wrong solutions, must score 0)	1,853 (1–3 per task)

Task types

type	tasks	the agent must
`fix_bug`	206	fix the described wrong behaviour in the named area; visible failing tests show it
`add_feature`	116	add a small documented feature (a function, an option, a field)
`locate_fix` 🔎	81	find where a described bug lives (the file is not named), then fix it
`edge_case`	75	handle a boundary the code misses (empty input, ties, inclusive ranges, …)
`implement_stub`	70	implement a stubbed function from its docstring / comments
`search_answer` 🔎	52	find a fact in the codebase and write it to `answer.txt` without modifying anything
`add_validation`	41	reject invalid input exactly as specified
`find_and_change` 🔎	39	find where a setting or behaviour is defined (live file, not the decoy) and change it
`add_test`	38	write a test that passes on the code and fails on hidden mutants
`read_answer`	38	read the code to answer a question in `answer.txt`
`fix_shell_script`	34	fix a shell script's behaviour
`add_cli_flag`	31	add a command-line flag with the given semantics
`fix_sql_query`	26	fix a query so it returns the specified rows
`rename_symbol` 🔎	25	rename a symbol everywhere it is used (definitions, imports, calls), nothing else
`handle_missing_file`	21	make the program degrade gracefully when a file is absent
`traceback_fix` 🔎	21	start from a traceback, find the cause (often not where it crashes), fix it
`config_change`	17	change a configuration value in the live config
`update_call_sites` 🔎	15	replace every call of a deprecated function
others	46	`write_transform_script`, `fix_build`, `refactor_extract`, `add_make_target`, `fix_regex`, `remove_dead_code`, `bump_version`

🔎 = search task (search = true in the metadata).

Examples:

py-air-quality--pm25-breakpoint (locate_fix, en): PM2.5 values between 35.5 and 55.4 µg/m³ get an air quality index that is far too high (test_unhealthy_for_sensitive fails: 45.0 should give 124). The index is a linear interpolation in the EPA breakpoint table; one row of the table the code uses is wrong (the category "unhealthy for sensitive groups" spans index 101-150). Find and fix it. Run python3 -m pytest -q.

js-i18n-messages--missing-translation (traceback_fix, en): The German checkout page crashes: (a node traceback ending in TypeError: Cannot read properties of undefined (reading 'replace') inside src/i18n.js). Fix the cause so that missing translations behave as documented. Run node --test test/*.test.js.

pl-host-inventory--answer-ansible-module (search_answer, it): Quale modulo genera l'inventario in formato INI per Ansible? Scrivi il suo percorso relativo a /workdir (per esempio lib/X/Y.pm) in /workdir/answer.txt senza modificare nessun file.

Browse them: instruction · task.toml · checks.

Splits

split	tasks	use
`train`	228	training
`new_train`	599	training: smaller tasks, about a third of them search tasks
`test_seen_types`	28	evaluation: task types that appear in training, on unseen codebases
`test_ood`	27	evaluation: task types that never appear in training
`new_test`	120	evaluation: unseen codebases, the same mix as `new_train`

The two training splits can be used together (827 tasks); the three test splits are held out (175 tasks). Splits are made by codebase: no codebase of a test split appears in a training split. The split of each task is in task.toml [metadata] split and in manifest.jsonl.

How a task is laid out

tasks/<codebase>--<slug>/
  task.toml              name = "small-software-harbor/<id>"; [metadata] codebase, language, domain, task_type,
                         instruction_language, difficulty, split, files_touched, oracle_steps, toolchains, search;
                         [environment] docker_image, allow_internet = false; [verifier] / [agent] timeouts
  instruction.md         the request: what is wrong or what to add, how to check
  environment/workdir/   the starting codebase (copied to /workdir)
  environment/setup.sh   modes (scripts executable)
  tests/                 test.sh -> swe_check.py on check.json; hidden tests h<i>/, mutants/ (add_test), texts/
  solution/              solve.sh (replays edits.json, the oracle as tool calls), attacks.json (self-check only)
manifest.jsonl           one row per task (id, codebase, language, domain, task_type, instruction_language,
                         difficulty, split, files_touched, oracle_steps, n_attacks, n_files, toolchains, search)
registry.json            the task index; build_report.json: deterministic build hashes
base_image/Dockerfile    the shared offline image
generators/              specs/ (the hand-written codebases and tasks as Python data), lib/ (DSL, verifiers),
                         build.py, check_local.py, check_docker.py

tests/ and solution/ are never visible to the agent.

Verified by tests, hardened by attacks

The verifier (tests/swe_check.py, python3 only) is all-or-nothing:

Collateral: protected files (the tests, unrelated sources) keep their sha256; no new source or config file may appear (no .bak copies, no helper modules smuggled next to the code); scratch files outside the source tree are allowed.
Behaviour: the visible project tests must pass, then the hidden tests run in a copy of the working directory, so nothing the agent leaves behind can tamper with them.
Answers: search / read tasks check answer.txt exactly (path or value) and require that no file changed.
Text checks: required / forbidden patterns where behaviour alone cannot tell (e.g. a rename left an alias).

Attacks. Each task carries 1–3 plausible wrong solutions, written for that task: the fix in the wrong file, an off-by-one variant, a partial feature, an alias left behind after a rename, the decoy config edited instead of the live one.

Self-check, every task, in Docker (generators/check_docker.py, on harbor-swe-base:2, no network):

mode	expected	result
oracle (`solve.sh`), oracle + scratch files	1	1 on every task
doing nothing	0	0 on every task
every attack (1,853)	0	0
delete the protected tests	0	0
oracle + a stray source file	0	0

All 1,002 tasks pass every mode.

Train with palingenesis

palingenesis runs these tasks natively: Docker containers per episode, the hidden tests as the reward, and emulated agent harnesses (opencode, pi, OpenHands, mini-swe-agent, …) so a policy learns more than one tool dialect. Install it (installation guide), validate the tasks and convert them into RL rows:

git clone https://github.com/mii-llm/palingenesis.git && cd palingenesis
uv sync                      # CPU is enough to check tasks and build rows; add --extra train for RL on a GPU
source .venv/bin/activate

hf download efederici/smol-software-harbor --repo-type dataset --local-dir smol-software-harbor
pgs harbor check smol-software-harbor/tasks               # every oracle must pass, doing nothing must fail
pgs harbor rows smol-software-harbor/tasks rows.jsonl     # convert the tasks into RL rows (task_dir + prompt)
pgs rl --config configs/rl_harbor.yaml      # multi-turn RL; env palingenesis.rl.envs.harbor:HarborEnv

Guides: Harbor environments · Harness profiles.

Run

Image. All tasks run in one shared image; task.toml names it harbor-swe-base:1 or harbor-swe-base:2, so build it once and give it both tags. Offline Debian bookworm-slim with python 3.11 + pytest, node 18, go 1.19, gcc 12 + make, ruby 3.1 + minitest, php 8.2, OpenJDK 17, rust 1.63, sqlite3, perl 5.36, lua 5.4, git, jq and ripgrep. Nothing is downloaded at task time (no pip / npm / cargo / go module fetches). Build it once:

hf download efederici/smol-software-harbor --repo-type dataset --local-dir smol-software-harbor
docker build -t harbor-swe-base:2 smol-software-harbor/base_image
docker tag harbor-swe-base:2 harbor-swe-base:1      # one image, both tags

Harbor CLI. The stock CLI does not stage environment/workdir/ or run setup.sh: convert the tasks to Harbor's single-step layout first (starting files under the step's workdir/).

Filter with datasets:

from datasets import load_dataset
m = load_dataset("efederici/smol-software-harbor", split="tasks")
python_search = m.filter(lambda r: r["language"] == "python" and r["search"] and r["split"] == "new_train")

Rebuild (deterministic, byte-identical output):

cd smol-software-harbor/generators && python3 build.py      # tasks/, manifest.jsonl, registry.json, build_report.json
python3 check_local.py                                      # local toolchains
python3 check_docker.py --workers 8                         # in Docker, every mode above

Known limits

Small by design: these tasks teach and measure basic agentic coding habits (read, locate, edit, test, stop), not large-repository engineering.
Attacks reduce reward hacking, they do not rule it out: a determined policy could still find a wrong-but-passing solution for some tasks.
Language coverage is uneven: Go, PHP and Java have few tasks (23, 9, 9). Hard tasks (level 3) are 7%.
No independent human reading audit yet: oracles and attacks are verified by execution.

License and provenance

Apache-2.0. The codebases, tasks, attacks and verifiers were written with Claude agents and built deterministically from hand-written specs: no scraped code, no personal data. Names, domains and data in the codebases are fictional.

Downloads last month: 101

Collection including efederici/smol-software-harbor

Harbor agent tasks

Collection

Verifiable Harbor tasks for training and evaluating small agents (shell, file editing, software engineering). Ready for RL with palingenesis. • 4 items • Updated 1 day ago