Datasets:

RedMod
/

recursive-v4-flash-rejected-samples-filtered-0.7thresh

prompt_id stringlengths 20 20	task_id stringlengths 33 33	dataset_index int64 0 29.5k	skip_bucket stringclasses 1 value	try int64 1 16	harness stringclasses 3 values	model stringclasses 2 values	prompt stringlengths 67 1.77k	docker_image stringlengths 52 52	cwd stringclasses 63 values	difficulty stringclasses 3 values	category stringclasses 42 values	n_turns int64 0 232	submitted bool 2 classes	reasons stringlengths 47 1.23k	reward stringclasses 10 values	test_sh_returncode int64 0 137 ⌀	test_sh_tail stringlengths 0 2k ⌀	tests_total int64 1 29 ⌀	tests_passed int64 0 28 ⌀	tests_failed int64 0 23 ⌀	tests_pass_pct float64 0 100 ⌀	messages listlengths 0 548	turns listlengths 1 232 ⌀	tools stringclasses 4 values	env_cost float64 0 0.16	token_cost float64 0 0.37	prompt_hint stringclasses 1 value	row_uuid stringlengths 36 36
000988dce5f1a79bad06	rts_task_a1f7a50adbe965e42492898f	18,370	failed_task	1	hermes	deepseek/deepseek-v4-flash	In the `/app` directory, a Judge0 API response is stored at `data/judge0_response.json` containing a base64-encoded program output in its `stdout` field. Extract and decode that output, writing the exact raw bytes to `result.txt`. Compute the SHA-256 hash of `result.txt` and store its hex digest in `integrity.txt`. The...	registry-b.deepgrove.ai/rts:a1f7a50adbe965e42492898f	/app	easy	scripting-automation	10	true	["tests failed: {\"reward\": \"0\", \"test_sh_returncode\": 0, \"test_sh_tail\": \"abcdef' for c in stripped), 'Integrity hash contains non-hex characters'\\n \\n # --- repair note ---\\n assert os.path.isfile(REPAIR_NOTE_PATH), f'{REPAIR_NOTE_PATH} does not exist'\\n with open(REPAIR_NOTE_PATH,...	0	0	abcdef' for c in stripped), 'Integrity hash contains non-hex characters' # --- repair note --- assert os.path.isfile(REPAIR_NOTE_PATH), f'{REPAIR_NOTE_PATH} does not exist' with open(REPAIR_NOTE_PATH, 'r') as f: note = f.read().strip() assert note, "repair_note.txt is em...	4	1	3	25	[ { "role": "system", "content": "You are Hermes Agent, an intelligent AI assistant created by Nous Research. You are helpful, knowledgeable, and direct. You assist users with a wide range of tasks including answering questions, writing and editing code, analyzing information, creative work, and executing act...	[ { "turn": 1, "reasoning": "Let me start by reading the necessary files to understand the task requirements and the data I'm working with.", "n_tool_calls": 3 }, { "turn": 2, "reasoning": "Let me enumerate all requirements:\n\n1. Requirement 1: Read `/app/data/judge0_response.json`, extra...	[{"name": "clarify", "description": "Ask the user a question when you need clarification, feedback, or a decision before proceeding. Supports three modes:\n\n1. Single-select multiple choice — provide up to 4 choices. The user picks one or types their own answer via a 5th 'Other' option.\n2. **Multi-select multiple...	0.004177	0.005757	Be extremely self-critical and grounded. Treat every belief about the state of the system as UNVERIFIED until a tool result proves it. Never assert that something exists, works, or is correct because you wrote it, intended it, or because it seems obvious -- only because a command you ran returned output that shows it. ...	ce018fa6-55f2-4be7-888b-c55fe74d3741
0016bbf50f93284c474e	rts_task_cdc8c7d72495a706d9bd1fa4	23,464	failed_task	1	hermes	deepseek/deepseek-v4-flash	"In /app, repair the KSQL validation pipeline.\nFix validate.sh to count only standalone semicolons (...TRUNCATED)	registry-b.deepgrove.ai/rts:cdc8c7d72495a706d9bd1fa4	/app	easy	database-operations	33	true	"[\"tests failed: {\\\"reward\\\": \\\"0\\\", \\\"test_sh_returncode\\\": 1, \\\"test_sh_tail\\\": \(...TRUNCATED)	0	1	"verify_out.read_text()\n assert \"Independent count:\" in vtext, \"verify_output.txt missing(...TRUNCATED)	4	2	2	50	[{"role":"system","content":"You are Hermes Agent, an intelligent AI assistant created by Nous Resea(...TRUNCATED)	[{"turn":1,"reasoning":"Let me start by enumerating every requirement from the task statement as a n(...TRUNCATED)	"[{\"name\": \"clarify\", \"description\": \"Ask the user a question when you need clarification, fe(...TRUNCATED)	0.006958	0.017836	"Be extremely self-critical and grounded. Treat every belief about the state of the system as UNVERI(...TRUNCATED)	1103681c-4bb4-4f43-91f1-3d0d62199e6b
0019a648390fb6d06c7c	rts_task_c36f6bea1af4bade36f6ff98	22,333	failed_task	1	pi	deepseek/deepseek-v4-flash	"Use the scripts and guide in /opt: gen_results.py (buggy generator), check.py (validator), and evid(...TRUNCATED)	registry-b.deepgrove.ai/rts:c36f6bea1af4bade36f6ff98	/app	easy	scripting-automation	30	true	"[\"tests failed: {\\\"reward\\\": \\\"0\\\", \\\"test_sh_returncode\\\": 0, \\\"test_sh_tail\\\": \(...TRUNCATED)	0	0	"ne.strip())\n ]\n fail_lines = [\n line.strip() for line in report.splitli(...TRUNCATED)	14	12	2	85.71	[{"role":"system","content":"You are an expert coding assistant operating inside pi, a coding agent (...TRUNCATED)	[{"turn":1,"reasoning":"Let me break down this task carefully:\n\n1. First, I need to read the files(...TRUNCATED)	"[{\"type\": \"function\", \"function\": {\"name\": \"bash\", \"description\": \"Execute a bash comm(...TRUNCATED)	0.004868	0.006925	"Be extremely self-critical and grounded. Treat every belief about the state of the system as UNVERI(...TRUNCATED)	f7c2a16c-39a6-418c-9393-a3c038f2ed4e
0021a8ebba6ea179d4e8	rts_task_d645877cf8076ba8b7bef425	24,510	failed_task	1	pi	deepseek/deepseek-v4-flash	"In /app, reconcile the Pacemaker/Corosync cluster init workflow: update `cluster-config.yaml` so al(...TRUNCATED)	registry-b.deepgrove.ai/rts:d645877cf8076ba8b7bef425	/app	easy	system-administration	17	true	"[\"tests failed: {\\\"reward\\\": \\\"MISSING\\\", \\\"test_sh_returncode\\\": 1, \\\"test_sh_tail\(...TRUNCATED)	MISSING	1	"PASS check_01_required_evidence\nFAIL check_02_intermediate_artifacts: result.txt line 3 mismatch: (...TRUNCATED)	null	null	null	null	[{"role":"system","content":"You are an expert coding assistant operating inside pi, a coding agent (...TRUNCATED)	[{"turn":1,"reasoning":"Let me start by understanding the full task by reading the relevant files. F(...TRUNCATED)	"[{\"type\": \"function\", \"function\": {\"name\": \"bash\", \"description\": \"Execute a bash comm(...TRUNCATED)	0.001915	0.005086	"Be extremely self-critical and grounded. Treat every belief about the state of the system as UNVERI(...TRUNCATED)	01f70510-5508-43fd-a72d-22f5af7f3d32
002d2eda86153d740f08	rts_task_3770a2a3cc97d0db14a2f1ff	6,179	failed_task	1	pi	deepseek/deepseek-v4-flash	"In /app, implement solution.py according to /app/spec.md, using argparse and an accumulator to prod(...TRUNCATED)	registry-b.deepgrove.ai/rts:3770a2a3cc97d0db14a2f1ff	/app	easy	scripting-automation	37	true	"[\"tests failed: {\\\"reward\\\": \\\"0\\\", \\\"test_sh_returncode\\\": 1, \\\"test_sh_tail\\\": \(...TRUNCATED)	0	1	"ected package curl.\r\nPreparing to unpack .../12-curl_7.81.0-1ubuntu1.26_amd64.deb ...\r\nUnpackin(...TRUNCATED)	null	null	null	null	[{"role":"system","content":"You are an expert coding assistant operating inside pi, a coding agent (...TRUNCATED)	[{"turn":1,"reasoning":"Let me start by reading the spec.md file to understand what needs to be impl(...TRUNCATED)	"[{\"type\": \"function\", \"function\": {\"name\": \"bash\", \"description\": \"Execute a bash comm(...TRUNCATED)	0.007465	0.027305	"Be extremely self-critical and grounded. Treat every belief about the state of the system as UNVERI(...TRUNCATED)	0f5bee77-88e2-4b68-b51d-9d8d2692ac72
00301d48eab5acd48e65	rts_task_91bfd696784632e23a98f996	16,411	failed_task	1	hermes	deepseek/deepseek-v4-flash	"Work in /app.\nThe session transcript is stored as session.txt somewhere under /app (not directly i(...TRUNCATED)	registry-b.deepgrove.ai/rts:91bfd696784632e23a98f996	/app	easy	data-processing	30	true	"[\"tests failed: {\\\"reward\\\": \\\"0\\\", \\\"test_sh_returncode\\\": 1, \\\"test_sh_tail\\\": \(...TRUNCATED)	0	1	"ms\n\n../tests/test_state.py ...F [100%]\n\n==========(...TRUNCATED)	4	3	1	75	[{"role":"system","content":"You are Hermes Agent, an intelligent AI assistant created by Nous Resea(...TRUNCATED)	[{"turn":1,"reasoning":"Let me start by understanding all the requirements in this task. Let me enum(...TRUNCATED)	"[{\"name\": \"clarify\", \"description\": \"Ask the user a question when you need clarification, fe(...TRUNCATED)	0.009557	0.023747	"Be extremely self-critical and grounded. Treat every belief about the state of the system as UNVERI(...TRUNCATED)	65e27de8-de54-408d-bf46-449b6c508255
00324bb2bca2b6ddbde2	rts_task_996795dbc4cc8591aa72437d	17,394	failed_task	1	hermes	deepseek/deepseek-v4-flash	"Repair /app/dev14.php to produce the correct outputs and complete workspace audit as documented in (...TRUNCATED)	registry-b.deepgrove.ai/rts:996795dbc4cc8591aa72437d	/app	easy	debugging-troubleshooting	17	true	"[\"tests failed: {\\\"reward\\\": \\\"0\\\", \\\"test_sh_returncode\\\": 1, \\\"test_sh_tail\\\": \(...TRUNCATED)	0	1	"non_empty_ws, \"workspace-audit.log must not be empty\"\n assert non_empty_ws[-1] == \"WORKS(...TRUNCATED)	4	3	1	75	[{"role":"system","content":"You are Hermes Agent, an intelligent AI assistant created by Nous Resea(...TRUNCATED)	[{"turn":1,"reasoning":"Let me start by understanding the task thoroughly. I need to:\n\n1. Run the (...TRUNCATED)	"[{\"name\": \"clarify\", \"description\": \"Ask the user a question when you need clarification, fe(...TRUNCATED)	0.004313	0.006416	"Be extremely self-critical and grounded. Treat every belief about the state of the system as UNVERI(...TRUNCATED)	3b00cfcb-5581-446f-a4db-00a9e13d0f55
003399f1b20fc10822dd	rts_task_01a793f41ee0f870f037bcba	148	failed_task	1	hermes	deepseek/deepseek-v4-flash	"Install salt-master and salt-minion on this Debian Stretch system.\nEnable and start both services (...TRUNCATED)	registry-b.deepgrove.ai/rts:01a793f41ee0f870f037bcba	/app	easy	system-administration	52	true	"[\"tests failed: {\\\"reward\\\": \\\"0\\\", \\\"test_sh_returncode\\\": 0, \\\"test_sh_tail\\\": \(...TRUNCATED)	0	0	"one, f\"Invalid mtime for {pf} in JSON manifest\"\n assert abs(mtime_epoch - actual_(...TRUNCATED)	10	7	3	70	[{"role":"system","content":"You are Hermes Agent, an intelligent AI assistant created by Nous Resea(...TRUNCATED)	[{"turn":1,"reasoning":"Let me start by carefully enumerating every requirement from the task statem(...TRUNCATED)	"[{\"name\": \"clarify\", \"description\": \"Ask the user a question when you need clarification, fe(...TRUNCATED)	0.011665	0.024294	"Be extremely self-critical and grounded. Treat every belief about the state of the system as UNVERI(...TRUNCATED)	91c3770b-b352-4476-ae31-48d95ce3277b
003b690eb38b616de077	rts_task_f2b4184d66df56e8a10e2082	27,677	failed_task	1	pi	deepseek/deepseek-v4-flash	"In the git repository at /app, correct the pre-existing LICENSE.txt so it properly credits the orig(...TRUNCATED)	registry-b.deepgrove.ai/rts:f2b4184d66df56e8a10e2082	/app	easy	version-control	12	true	"[\"tests failed: {\\\"reward\\\": \\\"0\\\", \\\"test_sh_returncode\\\": 0, \\\"test_sh_tail\\\": \(...TRUNCATED)	0	0	", \"your name\", \"your-name\"]:\n assert phrase.lower() not in content.lower(), f\"CHAN(...TRUNCATED)	4	2	2	50	[{"role":"system","content":"You are an expert coding assistant operating inside pi, a coding agent (...TRUNCATED)	[{"turn":1,"reasoning":"Let me first enumerate every requirement as a numbered checklist, then inves(...TRUNCATED)	"[{\"type\": \"function\", \"function\": {\"name\": \"bash\", \"description\": \"Execute a bash comm(...TRUNCATED)	0.00267	0.003337	"Be extremely self-critical and grounded. Treat every belief about the state of the system as UNVERI(...TRUNCATED)	6c462553-c0cc-497f-aed9-11364c80b80a
00613e8d55a0d864ef3f	rts_task_711782edd460de08f8ff0d06	12,806	failed_task	1	hermes	deepseek/deepseek-v4-flash	"extra-array is installed in /app.\nUse its in-place insertion sort to sort the array `[-2, -3, 1, 4(...TRUNCATED)	registry-b.deepgrove.ai/rts:711782edd460de08f8ff0d06	/app	easy	software-development	18	true	"[\"tests failed: {\\\"reward\\\": \\\"0\\\", \\\"test_sh_returncode\\\": 1, \\\"test_sh_tail\\\": \(...TRUNCATED)	0	1	"idence = False\n for root, dirs, files in os.walk(\"/app\"):\n # Skip node_module(...TRUNCATED)	4	3	1	75	[{"role":"system","content":"You are Hermes Agent, an intelligent AI assistant created by Nous Resea(...TRUNCATED)	[{"turn":1,"reasoning":"Let me carefully enumerate every requirement before doing any work.\n\n1. `e(...TRUNCATED)	"[{\"name\": \"clarify\", \"description\": \"Ask the user a question when you need clarification, fe(...TRUNCATED)	0.003836	0.005034	"Be extremely self-critical and grounded. Treat every belief about the state of the system as UNVERI(...TRUNCATED)	4cd896ba-e521-4d5c-9cd5-6be8e51480cc

End of preview. Expand in Data Studio

recursive-v4-flash-rejected-samples (filtered, 0.7 threshold)

A quality-filtered subset of RedMod/recursive-v4-flash-rejected-samples. These are deepseek-v4-flash trajectories that FAILED their automated tests, but were kept because an LLM grader judged the attempt substantively close to correct.

Filter

Each trace was graded (facts-extraction rubric) into per-trace counts. A trace is kept iff all hold:

artifacts_requested == artifacts_delivered (the agent produced every requested deliverable), and
artifacts_delivered > 0, and
artifacts_substantively_matching / artifacts_delivered >= 0.7.

46,731 / 86,470 traces (54.0%) pass. Each row carries its row_uuid (per-trace key). Grades were produced with openai/gpt-5.6-luna, which is a LENIENT grader relative to a stricter baseline (e.g. claude-haiku-4.5) — treat this subset as a high-recall triage set, not a verified-positive set.

Downloads last month: 35