Anthroly Research
Collection
RL environments, RLVR, reward hacking research by Anthroly (anthroly.com) • 2 items • Updated
task_id stringlengths 15 15 | category stringclasses 5
values | prompt stringlengths 65 153 | grader_type stringclasses 1
value | grader stringlengths 105 299 | reference_solution stringlengths 3 89 | difficulty stringclasses 2
values | provenance stringclasses 1
value |
|---|---|---|---|---|---|---|---|
anth-sample-001 | arithmetic | Compute 70 × 95 − 7. Put only the final integer on the last line. | python | def grade(response: str) -> float:
import re
m = re.findall(r"-?\d+(?:\.\d+)?", response.strip().splitlines()[-1] if response.strip() else "")
return 1.0 if m and abs(float(m[-1]) - 6643) < 1e-6 else 0.0 | 6643 | easy | synthetic: programmatically generated by Anthroly as an illustrative example; not expert-annotated |
anth-sample-002 | arithmetic | Compute 39 × 73 − 2. Put only the final integer on the last line. | python | def grade(response: str) -> float:
import re
m = re.findall(r"-?\d+(?:\.\d+)?", response.strip().splitlines()[-1] if response.strip() else "")
return 1.0 if m and abs(float(m[-1]) - 2845) < 1e-6 else 0.0 | 2845 | easy | synthetic: programmatically generated by Anthroly as an illustrative example; not expert-annotated |
anth-sample-003 | arithmetic | Compute 95 × 75 − 3. Put only the final integer on the last line. | python | def grade(response: str) -> float:
import re
m = re.findall(r"-?\d+(?:\.\d+)?", response.strip().splitlines()[-1] if response.strip() else "")
return 1.0 if m and abs(float(m[-1]) - 7122) < 1e-6 else 0.0 | 7122 | easy | synthetic: programmatically generated by Anthroly as an illustrative example; not expert-annotated |
anth-sample-004 | arithmetic | Compute 30 × 16 − 7. Put only the final integer on the last line. | python | def grade(response: str) -> float:
import re
m = re.findall(r"-?\d+(?:\.\d+)?", response.strip().splitlines()[-1] if response.strip() else "")
return 1.0 if m and abs(float(m[-1]) - 473) < 1e-6 else 0.0 | 473 | easy | synthetic: programmatically generated by Anthroly as an illustrative example; not expert-annotated |
anth-sample-005 | arithmetic | Compute 77 × 38 − 3. Put only the final integer on the last line. | python | def grade(response: str) -> float:
import re
m = re.findall(r"-?\d+(?:\.\d+)?", response.strip().splitlines()[-1] if response.strip() else "")
return 1.0 if m and abs(float(m[-1]) - 2923) < 1e-6 else 0.0 | 2923 | easy | synthetic: programmatically generated by Anthroly as an illustrative example; not expert-annotated |
anth-sample-006 | arithmetic | Compute 43 × 66 − 3. Put only the final integer on the last line. | python | def grade(response: str) -> float:
import re
m = re.findall(r"-?\d+(?:\.\d+)?", response.strip().splitlines()[-1] if response.strip() else "")
return 1.0 if m and abs(float(m[-1]) - 2835) < 1e-6 else 0.0 | 2835 | easy | synthetic: programmatically generated by Anthroly as an illustrative example; not expert-annotated |
anth-sample-007 | arithmetic | Compute 83 × 40 − 5. Put only the final integer on the last line. | python | def grade(response: str) -> float:
import re
m = re.findall(r"-?\d+(?:\.\d+)?", response.strip().splitlines()[-1] if response.strip() else "")
return 1.0 if m and abs(float(m[-1]) - 3315) < 1e-6 else 0.0 | 3315 | easy | synthetic: programmatically generated by Anthroly as an illustrative example; not expert-annotated |
anth-sample-008 | arithmetic | Compute 31 × 18 − 2. Put only the final integer on the last line. | python | def grade(response: str) -> float:
import re
m = re.findall(r"-?\d+(?:\.\d+)?", response.strip().splitlines()[-1] if response.strip() else "")
return 1.0 if m and abs(float(m[-1]) - 556) < 1e-6 else 0.0 | 556 | easy | synthetic: programmatically generated by Anthroly as an illustrative example; not expert-annotated |
anth-sample-009 | arithmetic | Compute 32 × 26 − 6. Put only the final integer on the last line. | python | def grade(response: str) -> float:
import re
m = re.findall(r"-?\d+(?:\.\d+)?", response.strip().splitlines()[-1] if response.strip() else "")
return 1.0 if m and abs(float(m[-1]) - 826) < 1e-6 else 0.0 | 826 | easy | synthetic: programmatically generated by Anthroly as an illustrative example; not expert-annotated |
anth-sample-010 | arithmetic | Compute 71 × 64 − 6. Put only the final integer on the last line. | python | def grade(response: str) -> float:
import re
m = re.findall(r"-?\d+(?:\.\d+)?", response.strip().splitlines()[-1] if response.strip() else "")
return 1.0 if m and abs(float(m[-1]) - 4538) < 1e-6 else 0.0 | 4538 | easy | synthetic: programmatically generated by Anthroly as an illustrative example; not expert-annotated |
anth-sample-011 | string-manipulation | Sort these words alphabetically and output them comma-separated with no spaces: omega lambda gamma alpha | python | def grade(response: str) -> float:
return 1.0 if response.strip() == 'alpha,gamma,lambda,omega' else 0.0 | alpha,gamma,lambda,omega | easy | synthetic: programmatically generated by Anthroly as an illustrative example; not expert-annotated |
anth-sample-012 | string-manipulation | Sort these words alphabetically and output them comma-separated with no spaces: omega gamma delta sigma | python | def grade(response: str) -> float:
return 1.0 if response.strip() == 'delta,gamma,omega,sigma' else 0.0 | delta,gamma,omega,sigma | easy | synthetic: programmatically generated by Anthroly as an illustrative example; not expert-annotated |
anth-sample-013 | string-manipulation | Sort these words alphabetically and output them comma-separated with no spaces: omega lambda beta zeta | python | def grade(response: str) -> float:
return 1.0 if response.strip() == 'beta,lambda,omega,zeta' else 0.0 | beta,lambda,omega,zeta | easy | synthetic: programmatically generated by Anthroly as an illustrative example; not expert-annotated |
anth-sample-014 | string-manipulation | Sort these words alphabetically and output them comma-separated with no spaces: kappa alpha sigma lambda | python | def grade(response: str) -> float:
return 1.0 if response.strip() == 'alpha,kappa,lambda,sigma' else 0.0 | alpha,kappa,lambda,sigma | easy | synthetic: programmatically generated by Anthroly as an illustrative example; not expert-annotated |
anth-sample-015 | string-manipulation | Sort these words alphabetically and output them comma-separated with no spaces: beta alpha gamma delta | python | def grade(response: str) -> float:
return 1.0 if response.strip() == 'alpha,beta,delta,gamma' else 0.0 | alpha,beta,delta,gamma | easy | synthetic: programmatically generated by Anthroly as an illustrative example; not expert-annotated |
anth-sample-016 | string-manipulation | Sort these words alphabetically and output them comma-separated with no spaces: kappa gamma omega beta | python | def grade(response: str) -> float:
return 1.0 if response.strip() == 'beta,gamma,kappa,omega' else 0.0 | beta,gamma,kappa,omega | easy | synthetic: programmatically generated by Anthroly as an illustrative example; not expert-annotated |
anth-sample-017 | string-manipulation | Sort these words alphabetically and output them comma-separated with no spaces: gamma theta omega delta | python | def grade(response: str) -> float:
return 1.0 if response.strip() == 'delta,gamma,omega,theta' else 0.0 | delta,gamma,omega,theta | easy | synthetic: programmatically generated by Anthroly as an illustrative example; not expert-annotated |
anth-sample-018 | string-manipulation | Sort these words alphabetically and output them comma-separated with no spaces: zeta theta beta sigma | python | def grade(response: str) -> float:
return 1.0 if response.strip() == 'beta,sigma,theta,zeta' else 0.0 | beta,sigma,theta,zeta | easy | synthetic: programmatically generated by Anthroly as an illustrative example; not expert-annotated |
anth-sample-019 | structured-output | Return a single JSON object with exactly these keys and no others: status, name, tags, owner. Values may be anything. Output only the JSON. | python | def grade(response: str) -> float:
import json
try:
obj = json.loads(response.strip())
except Exception:
return 0.0
if not isinstance(obj, dict):
return 0.0
return 1.0 if sorted(obj.keys()) == ['name', 'owner', 'status', 'tags'] else 0.0 | {"status": null, "name": null, "tags": null, "owner": null} | medium | synthetic: programmatically generated by Anthroly as an illustrative example; not expert-annotated |
anth-sample-020 | structured-output | Return a single JSON object with exactly these keys and no others: status, priority, name, score. Values may be anything. Output only the JSON. | python | def grade(response: str) -> float:
import json
try:
obj = json.loads(response.strip())
except Exception:
return 0.0
if not isinstance(obj, dict):
return 0.0
return 1.0 if sorted(obj.keys()) == ['name', 'priority', 'score', 'status'] else 0.0 | {"status": null, "priority": null, "name": null, "score": null} | medium | synthetic: programmatically generated by Anthroly as an illustrative example; not expert-annotated |
anth-sample-021 | structured-output | Return a single JSON object with exactly these keys and no others: score, tags, status, owner, name, id. Values may be anything. Output only the JSON. | python | def grade(response: str) -> float:
import json
try:
obj = json.loads(response.strip())
except Exception:
return 0.0
if not isinstance(obj, dict):
return 0.0
return 1.0 if sorted(obj.keys()) == ['id', 'name', 'owner', 'score', 'status', 'tags'] else 0.0 | {"score": null, "tags": null, "status": null, "owner": null, "name": null, "id": null} | medium | synthetic: programmatically generated by Anthroly as an illustrative example; not expert-annotated |
anth-sample-022 | structured-output | Return a single JSON object with exactly these keys and no others: id, priority, owner, name. Values may be anything. Output only the JSON. | python | def grade(response: str) -> float:
import json
try:
obj = json.loads(response.strip())
except Exception:
return 0.0
if not isinstance(obj, dict):
return 0.0
return 1.0 if sorted(obj.keys()) == ['id', 'name', 'owner', 'priority'] else 0.0 | {"id": null, "priority": null, "owner": null, "name": null} | medium | synthetic: programmatically generated by Anthroly as an illustrative example; not expert-annotated |
anth-sample-023 | structured-output | Return a single JSON object with exactly these keys and no others: status, priority, owner. Values may be anything. Output only the JSON. | python | def grade(response: str) -> float:
import json
try:
obj = json.loads(response.strip())
except Exception:
return 0.0
if not isinstance(obj, dict):
return 0.0
return 1.0 if sorted(obj.keys()) == ['owner', 'priority', 'status'] else 0.0 | {"status": null, "priority": null, "owner": null} | medium | synthetic: programmatically generated by Anthroly as an illustrative example; not expert-annotated |
anth-sample-024 | structured-output | Return a single JSON object with exactly these keys and no others: priority, score, name, id. Values may be anything. Output only the JSON. | python | def grade(response: str) -> float:
import json
try:
obj = json.loads(response.strip())
except Exception:
return 0.0
if not isinstance(obj, dict):
return 0.0
return 1.0 if sorted(obj.keys()) == ['id', 'name', 'priority', 'score'] else 0.0 | {"priority": null, "score": null, "name": null, "id": null} | medium | synthetic: programmatically generated by Anthroly as an illustrative example; not expert-annotated |
anth-sample-025 | structured-output | Return a single JSON object with exactly these keys and no others: status, priority, name. Values may be anything. Output only the JSON. | python | def grade(response: str) -> float:
import json
try:
obj = json.loads(response.strip())
except Exception:
return 0.0
if not isinstance(obj, dict):
return 0.0
return 1.0 if sorted(obj.keys()) == ['name', 'priority', 'status'] else 0.0 | {"status": null, "priority": null, "name": null} | medium | synthetic: programmatically generated by Anthroly as an illustrative example; not expert-annotated |
anth-sample-026 | structured-output | Return a single JSON object with exactly these keys and no others: owner, tags, priority, name, status, id. Values may be anything. Output only the JSON. | python | def grade(response: str) -> float:
import json
try:
obj = json.loads(response.strip())
except Exception:
return 0.0
if not isinstance(obj, dict):
return 0.0
return 1.0 if sorted(obj.keys()) == ['id', 'name', 'owner', 'priority', 'status', 'tags'] else 0.0 | {"owner": null, "tags": null, "priority": null, "name": null, "status": null, "id": null} | medium | synthetic: programmatically generated by Anthroly as an illustrative example; not expert-annotated |
anth-sample-027 | unit-conversion | Convert 421 kilometres to miles (1 mile = 1.609344 km). Round to 2 decimal places and end your answer with the number. | python | def grade(response: str) -> float:
import re
m = re.findall(r"-?\d+(?:\.\d+)?", response)
return 1.0 if m and abs(float(m[-1]) - 261.6) <= 0.01 else 0.0 | 261.6 | easy | synthetic: programmatically generated by Anthroly as an illustrative example; not expert-annotated |
anth-sample-028 | unit-conversion | Convert 363 kilometres to miles (1 mile = 1.609344 km). Round to 2 decimal places and end your answer with the number. | python | def grade(response: str) -> float:
import re
m = re.findall(r"-?\d+(?:\.\d+)?", response)
return 1.0 if m and abs(float(m[-1]) - 225.56) <= 0.01 else 0.0 | 225.56 | easy | synthetic: programmatically generated by Anthroly as an illustrative example; not expert-annotated |
anth-sample-029 | unit-conversion | Convert 352 kilometres to miles (1 mile = 1.609344 km). Round to 2 decimal places and end your answer with the number. | python | def grade(response: str) -> float:
import re
m = re.findall(r"-?\d+(?:\.\d+)?", response)
return 1.0 if m and abs(float(m[-1]) - 218.72) <= 0.01 else 0.0 | 218.72 | easy | synthetic: programmatically generated by Anthroly as an illustrative example; not expert-annotated |
anth-sample-030 | unit-conversion | Convert 192 kilometres to miles (1 mile = 1.609344 km). Round to 2 decimal places and end your answer with the number. | python | def grade(response: str) -> float:
import re
m = re.findall(r"-?\d+(?:\.\d+)?", response)
return 1.0 if m and abs(float(m[-1]) - 119.3) <= 0.01 else 0.0 | 119.3 | easy | synthetic: programmatically generated by Anthroly as an illustrative example; not expert-annotated |
anth-sample-031 | unit-conversion | Convert 145 kilometres to miles (1 mile = 1.609344 km). Round to 2 decimal places and end your answer with the number. | python | def grade(response: str) -> float:
import re
m = re.findall(r"-?\d+(?:\.\d+)?", response)
return 1.0 if m and abs(float(m[-1]) - 90.1) <= 0.01 else 0.0 | 90.1 | easy | synthetic: programmatically generated by Anthroly as an illustrative example; not expert-annotated |
anth-sample-032 | unit-conversion | Convert 135 kilometres to miles (1 mile = 1.609344 km). Round to 2 decimal places and end your answer with the number. | python | def grade(response: str) -> float:
import re
m = re.findall(r"-?\d+(?:\.\d+)?", response)
return 1.0 if m and abs(float(m[-1]) - 83.89) <= 0.01 else 0.0 | 83.89 | easy | synthetic: programmatically generated by Anthroly as an illustrative example; not expert-annotated |
anth-sample-033 | unit-conversion | Convert 346 kilometres to miles (1 mile = 1.609344 km). Round to 2 decimal places and end your answer with the number. | python | def grade(response: str) -> float:
import re
m = re.findall(r"-?\d+(?:\.\d+)?", response)
return 1.0 if m and abs(float(m[-1]) - 214.99) <= 0.01 else 0.0 | 214.99 | easy | synthetic: programmatically generated by Anthroly as an illustrative example; not expert-annotated |
anth-sample-034 | unit-conversion | Convert 232 kilometres to miles (1 mile = 1.609344 km). Round to 2 decimal places and end your answer with the number. | python | def grade(response: str) -> float:
import re
m = re.findall(r"-?\d+(?:\.\d+)?", response)
return 1.0 if m and abs(float(m[-1]) - 144.16) <= 0.01 else 0.0 | 144.16 | easy | synthetic: programmatically generated by Anthroly as an illustrative example; not expert-annotated |
anth-sample-035 | code-reasoning | List all prime numbers between 29 and 44 inclusive, in increasing order, separated by spaces, on the last line. | python | def grade(response: str) -> float:
import re
got = [int(x) for x in re.findall(r"\d+", response.strip().splitlines()[-1] if response.strip() else "")]
return 1.0 if got == [29, 31, 37, 41, 43] else 0.0 | 29 31 37 41 43 | medium | synthetic: programmatically generated by Anthroly as an illustrative example; not expert-annotated |
anth-sample-036 | code-reasoning | List all prime numbers between 35 and 45 inclusive, in increasing order, separated by spaces, on the last line. | python | def grade(response: str) -> float:
import re
got = [int(x) for x in re.findall(r"\d+", response.strip().splitlines()[-1] if response.strip() else "")]
return 1.0 if got == [37, 41, 43] else 0.0 | 37 41 43 | medium | synthetic: programmatically generated by Anthroly as an illustrative example; not expert-annotated |
anth-sample-037 | code-reasoning | List all prime numbers between 15 and 25 inclusive, in increasing order, separated by spaces, on the last line. | python | def grade(response: str) -> float:
import re
got = [int(x) for x in re.findall(r"\d+", response.strip().splitlines()[-1] if response.strip() else "")]
return 1.0 if got == [17, 19, 23] else 0.0 | 17 19 23 | medium | synthetic: programmatically generated by Anthroly as an illustrative example; not expert-annotated |
anth-sample-038 | code-reasoning | List all prime numbers between 22 and 31 inclusive, in increasing order, separated by spaces, on the last line. | python | def grade(response: str) -> float:
import re
got = [int(x) for x in re.findall(r"\d+", response.strip().splitlines()[-1] if response.strip() else "")]
return 1.0 if got == [23, 29, 31] else 0.0 | 23 29 31 | medium | synthetic: programmatically generated by Anthroly as an illustrative example; not expert-annotated |
anth-sample-039 | code-reasoning | List all prime numbers between 15 and 29 inclusive, in increasing order, separated by spaces, on the last line. | python | def grade(response: str) -> float:
import re
got = [int(x) for x in re.findall(r"\d+", response.strip().splitlines()[-1] if response.strip() else "")]
return 1.0 if got == [17, 19, 23, 29] else 0.0 | 17 19 23 29 | medium | synthetic: programmatically generated by Anthroly as an illustrative example; not expert-annotated |
anth-sample-040 | code-reasoning | List all prime numbers between 6 and 41 inclusive, in increasing order, separated by spaces, on the last line. | python | def grade(response: str) -> float:
import re
got = [int(x) for x in re.findall(r"\d+", response.strip().splitlines()[-1] if response.strip() else "")]
return 1.0 if got == [7, 11, 13, 17, 19, 23, 29, 31, 37, 41] else 0.0 | 7 11 13 17 19 23 29 31 37 41 | medium | synthetic: programmatically generated by Anthroly as an illustrative example; not expert-annotated |
Anthroly RL Environment Sample (Synthetic)
A small, synthetic, illustrative sample of 40 RL task specifications, each paired with a verifiable Python grader of the kind used in reinforcement learning with verifiable rewards (RLVR). It shows the format Anthroly uses to package tasks and graders. It is not a production training set and was not expert-annotated: every task was generated programmatically, and every grader was checked to return 1.0 on the reference solution and 0.0 on a non-answer.
Built by Anthroly, which builds RL environments and datasets for AI labs. Maintainer: Sarath Kavuru.
Fields
| field | description |
|---|---|
task_id |
stable id |
category |
arithmetic, string-manipulation, structured-output, unit-conversion, code-reasoning |
prompt |
instruction given to the policy |
grader_type |
always python here |
grader |
source of grade(response: str) -> float returning 0.0 or 1.0 |
reference_solution |
a response that the grader accepts |
difficulty |
rough heuristic label (easy/medium); not calibrated against any model |
provenance |
generation note |
Usage
from datasets import load_dataset
ds = load_dataset("sarathkavuru/anthroly-rl-env-sample", split="train")
ns = {}
exec(ds[0]["grader"], ns) # graders are plain Python; review before executing, and sandbox untrusted code
print(ns["grade"](ds[0]["reference_solution"])) # 1.0
Limitations
- Tasks are easy and synthetic; most frontier models will solve them, so pass rates will be near 100% and offer little RL signal on their own (see pass-rate calibration).
- Graders are deliberately simple and can be gamed (for example, by emitting a number without reasoning). That is the point of the example: see reward hacking and What makes an RL environment training-grade?.
Related
- Research papers and article mirrors: sarathkavuru/anthroly-research
- RLVR vs RLHF
- Eval sample size calculator
License
CC BY 4.0. Please attribute "Anthroly (anthroly.com)".
- Downloads last month
- 50