turintech/GLM-5.2-MaxMin108-GGUF
Text Generation • 325B • Updated • 769 • 4
Datasets:
Dataset Preview
The full dataset viewer is not available (click to read why). Only showing a preview of the rows.
The dataset generation failed
Error code: DatasetGenerationError
Exception: ArrowNotImplementedError
Message: Cannot write struct type 'group_subtasks' with no child field to Parquet. Consider adding a dummy child field.
Traceback: Traceback (most recent call last):
File "/usr/local/lib/python3.14/site-packages/datasets/builder.py", line 1869, in _prepare_split_single
num_examples, num_bytes = writer.finalize()
~~~~~~~~~~~~~~~^^
File "/usr/local/lib/python3.14/site-packages/datasets/arrow_writer.py", line 781, in finalize
self.write_rows_on_file()
~~~~~~~~~~~~~~~~~~~~~~~^^
File "/usr/local/lib/python3.14/site-packages/datasets/arrow_writer.py", line 663, in write_rows_on_file
self._write_table(table)
~~~~~~~~~~~~~~~~~^^^^^^^
File "/usr/local/lib/python3.14/site-packages/datasets/arrow_writer.py", line 771, in _write_table
self._build_writer(inferred_schema=pa_table.schema)
~~~~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/usr/local/lib/python3.14/site-packages/datasets/arrow_writer.py", line 812, in _build_writer
self.pa_writer = pq.ParquetWriter(
~~~~~~~~~~~~~~~~^
self.stream,
^^^^^^^^^^^^
...<9 lines>...
},
^^
)
^
File "/usr/local/lib/python3.14/site-packages/pyarrow/parquet/core.py", line 1070, in __init__
self.writer = _parquet.ParquetWriter(
~~~~~~~~~~~~~~~~~~~~~~^
sink, schema,
^^^^^^^^^^^^^
...<18 lines>...
store_decimal_as_integer=store_decimal_as_integer,
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
**options)
^^^^^^^^^^
File "pyarrow/_parquet.pyx", line 2363, in pyarrow._parquet.ParquetWriter.__cinit__
File "pyarrow/error.pxi", line 155, in pyarrow.lib.pyarrow_internal_check_status
return check_status(status)
File "pyarrow/error.pxi", line 92, in pyarrow.lib.check_status
raise convert_status(status)
pyarrow.lib.ArrowNotImplementedError: Cannot write struct type 'group_subtasks' with no child field to Parquet. Consider adding a dummy child field.
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 1369, 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 948, 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 1694, 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 1880, in _prepare_split_single
raise DatasetGenerationError("An error occurred while generating the dataset") from e
datasets.exceptions.DatasetGenerationError: An error occurred while generating the datasetNeed help to make the dataset viewer work? Make sure to review how to configure the dataset viewer, and open a discussion for direct support.
results dict | group_subtasks dict | configs dict | versions dict | n-shot dict | higher_is_better dict | n-samples dict | config dict | git_hash string | date float64 | pretty_env_info string | transformers_version string | lm_eval_version string | upper_git_hash null | tokenizer_pad_token list | tokenizer_eos_token list | tokenizer_bos_token list | eot_token_id int64 | max_length int64 | task_hashes dict | model_source string | model_name string | model_name_sanitized string | system_instruction null | system_instruction_sha null | fewshot_as_multiturn null | chat_template null | chat_template_sha null | total_evaluation_time_seconds string |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
{
"ifeval": {
"name": "ifeval",
"alias": "ifeval",
"sample_len": 25,
"prompt_level_strict_acc,none": 0.32,
"prompt_level_strict_acc_stderr,none": 0.09521904571390466,
"inst_level_strict_acc,none": 0.40540540540540543,
"inst_level_strict_acc_stderr,none": "N/A",
"prompt_level_loose_acc,... | {} | {
"ifeval": {
"task": "ifeval",
"dataset_path": "google/IFEval",
"test_split": "train",
"doc_to_text": "prompt",
"doc_to_target": 0,
"unsafe_code": false,
"process_results": "def process_results(doc, results):\n inp = InputExample(\n key=doc[\"key\"],\n instruction_id_list... | {
"ifeval": 4
} | {
"ifeval": 0
} | {
"ifeval": {
"prompt_level_strict_acc": true,
"inst_level_strict_acc": true,
"prompt_level_loose_acc": true,
"inst_level_loose_acc": true
}
} | {
"ifeval": {
"original": 541,
"effective": 25
}
} | {
"model": "local-completions",
"model_args": {
"model": "glm-eval",
"tokenizer": "zai-org/GLM-5.2",
"base_url": "http://127.0.0.1:8090/v1/completions",
"num_concurrent": 16,
"max_retries": 1,
"tokenized_requests": false
},
"batch_size": 1,
"batch_sizes": [],
"device": "cuda:0",
"u... | gguf-v0.19.0-959-g7f575c39d | 1,786,729,412.053335 | PyTorch version: 2.13.0+cu130
Is debug build: False
CUDA used to build PyTorch: 13.0
ROCM used to build PyTorch: N/A
OS: Ubuntu 24.04.4 LTS (x86_64)
GCC version: (Ubuntu 12.4.0-2ubuntu1~24.04.1) 12.4.0
Clang version: Could not collect
CMake version: version 3.28.3
Libc version: glibc-2.39
Python version: 3.11.15 (mai... | 5.15.0 | 0.4.12 | null | [
"<|endoftext|>",
"154820"
] | [
"<|endoftext|>",
"154820"
] | [
"None"
] | 154,820 | 2,047 | {} | local-completions | glm-eval | glm-eval | null | null | null | null | null | 226.64011203497648 |
GLM-5.2 expert saliency and keep-lists
Everything needed to reproduce turintech/GLM-5.2-MaxMin108-GGUF from unsloth/GLM-5.2-GGUF: about fifteen minutes of disk I/O, no GPU, no calibration run.
Contents
keep_lists/: which experts survive, per layer.keep_mix108_maxmin.jsonproduces the published model. The others are the ablations (code-only, general-only and random selection at several K), four of which degenerate on out-of-domain prompts.saliency_merged/: per-expert saliency using the REAP criterion, gate weight times the norm of the expert output, accumulated over 1.93 M tokens. One file per domain across ten domains, plus the pooled total. 75 MoE layers by 256 experts, withcnt,gateandsalper expert.weights_maxmin.json: the max-min fair blend weights.corpus_manifest.json: sha256-verified composition of the calibration corpus.results/: rawlm-evaluation-harnessoutput for every benchmark in the model card.extras/: the importance matrix used for the-fastvariant (regenerating it takes about 32 minutes of GPU time), and the trained routers from the router-repair experiment described below.
Rebuilding the model
git clone https://github.com/Dorijan10/glm-expert-pruning
python tools/glm_prune_gguf.py \
--src GLM-5.2-UD-IQ2_M-merged.gguf \
--dst MaxMin108.gguf \
--keeplist keep_lists/keep_mix108_maxmin.json \
--drop-blocks 78
Surviving experts are copied byte for byte, and nothing is requantised.
Why the saliency data may be useful on its own
Across all ten domains, no expert is ever idle: mean dead-expert count is 0.0 per layer. Pruning always removes actively-used capacity. Code and literary prose share essentially zero saliency ordering (Spearman near 0), and the union of per-domain top-108 lists spans 238 of 256 experts, which is why single-corpus selection breaks out-of-domain generation.
extras/routers_trained_v2.tgz holds the output of a router-repair experiment that failed
informatively. Retraining the router against the parent's MoE output improved layer-local
reconstruction error by 9.90% and made real perplexity 2.61% worse, degrading monotonically
from the very first step of interpolation between the original and trained weights. The
verbatim-sliced router appears to sit at a local optimum of perplexity.
Method, evaluation and known costs: see the model card.
- Downloads last month
- 100
Models trained or fine-tuned on turintech/GLM-5.2-MaxMin108-saliency