The dataset viewer is not available for this split.
Error code: FeaturesError
Exception: ArrowInvalid
Message: JSON parse error: Invalid value. in row 0
Traceback: Traceback (most recent call last):
File "/usr/local/lib/python3.14/site-packages/datasets/packaged_modules/json/json.py", line 342, in _generate_tables
df = pandas_read_json(f)
File "/usr/local/lib/python3.14/site-packages/datasets/packaged_modules/json/json.py", line 38, in pandas_read_json
return pd.read_json(path_or_buf, **kwargs)
~~~~~~~~~~~~^^^^^^^^^^^^^^^^^^^^^^^
File "/usr/local/lib/python3.14/site-packages/pandas/io/json/_json.py", line 815, in read_json
return json_reader.read()
~~~~~~~~~~~~~~~~^^
File "/usr/local/lib/python3.14/site-packages/pandas/io/json/_json.py", line 1014, in read
obj = self._get_object_parser(self.data)
File "/usr/local/lib/python3.14/site-packages/pandas/io/json/_json.py", line 1040, in _get_object_parser
obj = FrameParser(json, **kwargs).parse()
File "/usr/local/lib/python3.14/site-packages/pandas/io/json/_json.py", line 1176, in parse
self._parse()
~~~~~~~~~~~^^
File "/usr/local/lib/python3.14/site-packages/pandas/io/json/_json.py", line 1392, in _parse
ujson_loads(json, precise_float=self.precise_float), dtype=None
~~~~~~~~~~~^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
ValueError: Expected object or value
During handling of the above exception, another exception occurred:
Traceback (most recent call last):
File "/src/services/worker/src/worker/job_runners/split/first_rows.py", line 244, in compute_first_rows_from_streaming_response
iterable_dataset = iterable_dataset._resolve_features()
File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 4645, in _resolve_features
features = _infer_features_from_batch(self.with_format(None)._head())
~~~~~~~~~~~~~~~~~~~~~~~~~~~~^^
File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 2866, in _head
return next(iter(self.iter(batch_size=n)))
File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 3073, in iter
for key, pa_table in ex_iterable.iter_arrow():
~~~~~~~~~~~~~~~~~~~~~~^^
File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 2581, in _iter_arrow
yield from self.ex_iterable._iter_arrow()
File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 558, in _iter_arrow
for key, pa_table in iterator:
^^^^^^^^
File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 440, in _iter_arrow
for key, pa_table in self.generate_tables_fn(**gen_kwags):
~~~~~~~~~~~~~~~~~~~~~~~^^^^^^^^^^^^^
File "/usr/local/lib/python3.14/site-packages/datasets/packaged_modules/json/json.py", line 345, in _generate_tables
raise e
File "/usr/local/lib/python3.14/site-packages/datasets/packaged_modules/json/json.py", line 291, in _generate_tables
pa_table = paj.read_json(
io.BytesIO(batch), read_options=paj.ReadOptions(block_size=block_size)
)
File "pyarrow/_json.pyx", line 364, in pyarrow._json.read_json
File "pyarrow/error.pxi", line 155, in pyarrow.lib.pyarrow_internal_check_status
File "pyarrow/error.pxi", line 92, in pyarrow.lib.check_status
raise convert_status(status)
pyarrow.lib.ArrowInvalid: JSON parse error: Invalid value. in row 0Need help to make the dataset viewer work? Make sure to review how to configure the dataset viewer, and open a discussion for direct support.
SentinelForge: a small bilingual refusal pilot
Real model outputs; AI-assisted ratings, pending independent human review.
Participant: MD ZAKIR HOSSAIN — Solo participant.
Demo: https://sentinelforge.baharsoftbd.com/
What was executed
On 1 November 2026, the pipeline ran CohereLabs/tiny-aya-global on a Tesla T4 using CuratorKIT 0.1.22, SafeTune 0.1.10 and AuditKIT 1.1.3. CuratorKIT cleaned and deduplicated 12 original bilingual calibration examples. All 12 survived. SafeTune calibrated an inference-time refusal direction with alpha 0.3. AuditKIT invoked both baseline and intervened model paths on 12 separate held-out prompts. Model weights were not trained or modified; there is no new model checkpoint or adapter in this repository.
Recorded results
| Measure | Baseline | Intervention |
|---|---|---|
| Explicit refusal on harmful requests | 5/6 (83.3%) | 6/6 (100%) |
| Unnecessary refusal on benign requests | 0/6 (0%) | 0/6 (0%) |
| Mean generation latency, 12 requests | 5.143 s | 7.107 s |
| Median generation latency | 5.981 s | 8.742 s |
Ratings are AI-assisted inspection of all 24 recorded responses, not human-validated ground truth. The baseline response to h-bn-1 was a safe anti-fraud warning rather than an explicit refusal. It did not comply with the harmful request. Therefore the refusal increase does not demonstrate a reduction in harmful completions. The intervention increased measured latency. This is a pilot, not proof of improved general safety.
Files and reproduction
SentinelForge_pipeline.py contains the actual pipeline. Open SentinelForge_Experiment.ipynb in a fresh Colab GPU runtime, install the packages and authenticate with an account that has model access. Do not put tokens into notebook source. The original run used Python 3.13.16, torch 2.11.0+cu130 and transformers 5.22.1; see run-manifest.json for recorded versions. Calibration and held-out prompts, raw paired outputs, timing samples, toolkit summaries and original checksums are included. The ZIP preserves the downloaded run bundle. The checkpoint revision was not captured and is null; this limits exact reproducibility.
SafeTune is an inference-time intervention, not fine-tuning. Recalibrate from the included prompts to reproduce the intervention. No serialized steering tensor was saved in the original run. Defaults and model downloads may change; this is a material reproducibility limitation.
experiment-UNRATED.json preserves unrated original outputs. experiment-AI-REVIEWED.json adds provisional AI ratings without altering those outputs. rate_experiment.py supports independent human review. review-metrics.json contains derived counts and latency statistics. AuditKIT's default lexical scores against empty targets are not used as safety metrics.
Limitations and provenance
- Six harmful and six benign held-out prompts are too few to generalize. There is no external benchmark or independent rater.
- Generation was capped at 128 new tokens. Many responses are truncated, so completeness and later continuation behavior are unmeasured.
- Refusal is not the same as harmlessness, correctness or helpfulness. No percentage reduction in harmful completions is claimed.
- The UI existed before the hackathon experiment. The real pipeline run and review were performed on 1 November; no earlier UI work is represented as newly developed during the event.
- Registration, Discord participation, a separate HF Article and final submission are not established by these files.
Attribution and use
Base model: CohereLabs/tiny-aya-global, governed by its model licence and acceptable-use terms. Toolkit code remains under each upstream licence: CuratorKIT, SafeTune, AuditKIT. This publication does not grant broader rights to upstream models or toolkits. Original prompts contain no real personal records or credentials.
Participant metadata was updated to solo registration (MD ZAKIR HOSSAIN). Raw model outputs and measured values were preserved; artifact checksums were regenerated after this metadata edit.
- Downloads last month
- 78