Datasets:

zichuanxu
/

closing-the-loop-generations

Error code:   DatasetGenerationError
Exception:    CastError
Message:      Couldn't cast
ckpt: string
data: string
offset: int64
problems: int64
n: int64
protocol: string
sampling: struct<temperature: double, top_p: double, top_k: int64, presence_penalty: double>
  child 0, temperature: double
  child 1, top_p: double
  child 2, top_k: int64
  child 3, presence_penalty: double
thinking: bool
max_tokens: int64
seed: int64
acc: double
cut_share: double
mean_tokens: double
tokens_sum: int64
gen_seconds: double
load_seconds: double
tok_per_s: double
cuda_visible: null
gpu_after_load: string
gpu_end: string
text: string
closed_think: bool
n_tokens: int64
finish: string
sample: int64
correct: bool
id: string
pred: string
to
{'id': Value('string'), 'sample': Value('int64'), 'correct': Value('bool'), 'pred': Value('string'), 'finish': Value('string'), 'n_tokens': Value('int64'), 'closed_think': Value('bool'), 'text': Value('string')}
because column names don't match
Traceback:    Traceback (most recent call last):
                File "/usr/local/lib/python3.14/site-packages/datasets/builder.py", line 1835, in _prepare_split_single
                  for key, table in generator:
                                    ^^^^^^^^^
                File "/src/services/worker/src/worker/job_runners/config/parquet_and_info.py", line 621, in wrapped
                  for item in generator(*args, **kwargs):
                              ~~~~~~~~~^^^^^^^^^^^^^^^^^
                File "/usr/local/lib/python3.14/site-packages/datasets/packaged_modules/json/json.py", line 361, in _generate_tables
                  self._cast_table(pa_table, json_field_paths=json_field_paths),
                  ~~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
                File "/usr/local/lib/python3.14/site-packages/datasets/packaged_modules/json/json.py", line 132, in _cast_table
                  pa_table = table_cast(pa_table, self.info.features.arrow_schema)
                File "/usr/local/lib/python3.14/site-packages/datasets/table.py", line 2381, in table_cast
                  return cast_table_to_schema(table, schema)
                File "/usr/local/lib/python3.14/site-packages/datasets/table.py", line 2309, in cast_table_to_schema
                  raise CastError(
                  ...<3 lines>...
                  )
              datasets.table.CastError: Couldn't cast
              ckpt: string
              data: string
              offset: int64
              problems: int64
              n: int64
              protocol: string
              sampling: struct<temperature: double, top_p: double, top_k: int64, presence_penalty: double>
                child 0, temperature: double
                child 1, top_p: double
                child 2, top_k: int64
                child 3, presence_penalty: double
              thinking: bool
              max_tokens: int64
              seed: int64
              acc: double
              cut_share: double
              mean_tokens: double
              tokens_sum: int64
              gen_seconds: double
              load_seconds: double
              tok_per_s: double
              cuda_visible: null
              gpu_after_load: string
              gpu_end: string
              text: string
              closed_think: bool
              n_tokens: int64
              finish: string
              sample: int64
              correct: bool
              id: string
              pred: string
              to
              {'id': Value('string'), 'sample': Value('int64'), 'correct': Value('bool'), 'pred': Value('string'), 'finish': Value('string'), 'n_tokens': Value('int64'), 'closed_think': Value('bool'), 'text': Value('string')}
              because column names don't match
              
              The above exception was the direct cause of the following exception:
              
              Traceback (most recent call last):
                File "/src/services/worker/src/worker/job_runners/config/parquet_and_info.py", line 1379, in compute_config_parquet_and_info_response
                  parquet_operations, partial, estimated_dataset_info = stream_convert_to_parquet(
                                                                        ~~~~~~~~~~~~~~~~~~~~~~~~~^
                      builder, max_dataset_size_bytes=max_dataset_size_bytes
                      ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
                  )
                  ^
                File "/src/services/worker/src/worker/job_runners/config/parquet_and_info.py", line 956, in stream_convert_to_parquet
                  builder._prepare_split(split_generator=splits_generators[split], file_format="parquet")
                  ~~~~~~~~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
                File "/usr/local/lib/python3.14/site-packages/datasets/builder.py", line 1702, in _prepare_split
                  for job_id, done, content in self._prepare_split_single(
                                               ~~~~~~~~~~~~~~~~~~~~~~~~~~^
                      gen_kwargs=gen_kwargs, job_id=job_id, **_prepare_split_args
                      ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
                  ):
                  ^
                File "/usr/local/lib/python3.14/site-packages/datasets/builder.py", line 1887, in _prepare_split_single
                  raise DatasetGenerationError("An error occurred while generating the dataset") from e
              datasets.exceptions.DatasetGenerationError: An error occurred while generating the dataset

Need help to make the dataset viewer work? Make sure to review how to configure the dataset viewer, and open a discussion for direct support.

id string	correct bool	pred string	finish string	n_tokens int64	closed_think bool	text string
math_train_precalculus_299	false	null	length	32,768	false	The problem asks for the value of $k$ such that the vector $\mathbf{v} = k \mathbf{a} + \frac{3}{4} \mathbf{b}$ is collinear with the line passing through vectors $\mathbf{a}$ and $\mathbf{b}$. Let the line be denoted by $L$. The direction of any vector on the line $L$ can be represented as a scalar multiple of some no...
math_train_precalculus_58	false	null	length	32,768	false	Here is my thought process for solving the determinant equation: 1. Analyze the Request: * The user wants to solve for $x$ in a specific determinant equation. * The determinant is given by $\begin{vmatrix} x + a & x & x \\ x & x + a & x \\ x & x & x + a \end{vmatrix} = 0$. * There's a condition ...
gsm8k_train_5524	true	200	stop	3,340	true	Here's my thought process for solving this word problem: 1. Analyze the Request: * Subject: A technician named Gondor. * Income Sources: Repairing phones and repairing laptops. * Prices/Weights: * Phone = $10 * Laptop = $20 * Data Points (by day): ...
gsm8k_train_500	true	30	stop	6,568	true	Here's a thinking process that leads to the solution: 1. Analyze the Request: * Characters: Joe, Derek, Tammy. * Action: Catching a ball (count = catches). * Given Information: * Total catches by Joe ($J$) = 23. * Relation between Joe and Derek: "Derek made four l...
gsm8k_train_7033	false	null	length	32,768	false	"Here's a thinking process that leads to the solution:\n\n1. Analyze the Problem:\n * **Go(...TRUNCATED)
math_train_geometry_306	false	null	length	32,768	false	"Here's my thought process for solving this problem:\n\n1. Analyze the Request:\n * **Give(...TRUNCATED)
math_train_algebra_817	true	2	stop	13,547	true	"Here is my thought process for solving the equation $27 = 3(9)^{x-1}$:\n\n1. **Analyze the Request(...TRUNCATED)
gsm8k_train_5282	false	null	length	32,768	false	"Here's a thinking process that leads to the solution:\n\n1. Analyze the Problem:\n * Init(...TRUNCATED)
math_train_number_theory_423	false	null	length	32,768	false	"Here's a thinking process that leads to the solution:\n\n1. Analyze the Problem:\n * **Qu(...TRUNCATED)
gsm8k_train_1497	false	null	length	32,768	false	"Here's a thinking process that leads to the solution:\n\n1. Analyze the Problem:\n * Init(...TRUNCATED)

End of preview.

Closing the Loop: all generations

Every model answer behind Closing the Loop: A Simple Rule Beats an Agent-Written Weight Edit for Reasoning Loops. Qwen3.5-0.8B with and without weight edits, on MATH-500, GSM8K test and practice sets. The edits include the one written by the LLM agent GPT-6 Astra (Takehi et al., 2026), called the Astra edit below. Code, problem sets, protocol and analysis: github.com/zichuanxu/closing-the-loop. Edited model: zichuanxu/Qwen3.5-0.8B-CLC-sel.

Layout

One folder per run, named as in the study record. Each holds:

records.jsonl.gz, one line per answer: id (problem id in study/data/*.jsonl of the GitHub repo), sample, correct (Math-Verify 0.9.0 on the last \boxed{}), pred, finish (stop or length), n_tokens, closed_think, text (the full generation).
summary.json: checkpoint, data, protocol and sampling, seed, token limit, accuracy (acc), truncated share (cut_share), mean tokens, throughput and the A100 used.

Main runs

Run	Model	Benchmark	Seeds
`m500_A_base`	base, thinking	MATH-500	0
`m500_E_nothink`	base, non-thinking	MATH-500	0
`m500_B_astra`, `_s1`, `_s2`	Astra edit	MATH-500	0, 1, 2
`m500_I_clc`	CLC	MATH-500	0
`m500_I2_clcsel`, `_s1`, `_s2`	CLC-sel	MATH-500	0, 1, 2
`m500_I3_auto`	automatic edit, layer 13	MATH-500	0
`gsm_L_astra`, `_s1`, `_s2`	Astra edit	GSM8K test	0, 1, 2
`gsm_K_clc`	CLC	GSM8K test	0
`gsm_K2_clcsel`, `_s1`, `_s2`	CLC-sel	GSM8K test	0, 1, 2
`gsm_K3_auto`	automatic edit, layer 13	GSM8K test	0
`m500_A_bf`, `m500_B_bf`, `m500_I2_bf`	base, Astra, CLC-sel with budget forcing	MATH-500	0
`m500_P_rp`	base with repetition penalty 1.1	MATH-500	0

Other prefixes: gen_extract_base (the 512 practice answers the edit is built from), dev_* (development set), res_* (reserve set used to choose the neuron rule and the repetition penalty), scr_* (layer screen), m500_C/D/F/G/H/J_* (other edits and non-thinking runs), smoke_* (pipeline checks).

All runs use thinking mode unless named nothink, with temperature 0.6, top-p 0.95, top-k 20, presence penalty 1.5 and 32,768 tokens (non-thinking: temperature 0.7, top-p 0.8, 8,192 tokens; gen_extract_base: 16,384 tokens). Budget-forcing runs (*_bf) close the thinking at 30,720 tokens.

License

MIT. The problems come from MATH and GSM8K under their own licenses; the texts were generated by Qwen3.5-0.8B (Apache 2.0) and edits of it.

Downloads last month: -

Total file size:

175 MB