Relapse: runs, logs and probe sets
Collection
Per-question results, AuditKIT runs, CircuitKIT scores and CuratorKIT-gated probe sets. • 2 items • Updated
Dataset Viewer
The dataset viewer is not available for this subset.
Cannot get the split names for the config 'default' of the dataset.
Exception: SplitsNotFoundError
Message: The split names could not be parsed from the dataset config.
Traceback: Traceback (most recent call last):
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
return check_status(status)
File "pyarrow/error.pxi", line 92, in pyarrow.lib.check_status
raise convert_status(status)
pyarrow.lib.ArrowInvalid: JSON parse error: Column() changed from object to string in row 0
During handling of the above exception, another exception occurred:
Traceback (most recent call last):
File "/usr/local/lib/python3.14/site-packages/datasets/inspect.py", line 286, in get_dataset_config_info
for split_generator in builder._split_generators(
~~~~~~~~~~~~~~~~~~~~~~~~~^
StreamingDownloadManager(base_path=builder.base_path, download_config=download_config)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
)
^
File "/usr/local/lib/python3.14/site-packages/datasets/packaged_modules/json/json.py", line 101, in _split_generators
pa_table = next(iter(self._generate_tables(**splits[0].gen_kwargs, allow_full_read=False)))[1]
~~~~^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/usr/local/lib/python3.14/site-packages/datasets/packaged_modules/json/json.py", line 306, in _generate_tables
batch = json_encode_fields_in_json_lines(original_batch, json_field_paths)
File "/usr/local/lib/python3.14/site-packages/datasets/utils/json.py", line 111, in json_encode_fields_in_json_lines
examples = [ujson_loads(line) for line in original_batch.splitlines()]
~~~~~~~~~~~^^^^^^
File "/usr/local/lib/python3.14/site-packages/datasets/utils/json.py", line 20, in ujson_loads
return pd.io.json.ujson_loads(*args, **kwargs)
~~~~~~~~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^
ValueError: Expected object or value
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/split_names.py", line 68, in compute_split_names_from_streaming_response
for split in get_dataset_split_names(
~~~~~~~~~~~~~~~~~~~~~~~^
path=dataset,
^^^^^^^^^^^^^
config_name=config,
^^^^^^^^^^^^^^^^^^^
token=hf_token,
^^^^^^^^^^^^^^^
)
^
File "/usr/local/lib/python3.14/site-packages/datasets/inspect.py", line 340, in get_dataset_split_names
info = get_dataset_config_info(
path,
...<6 lines>...
**config_kwargs,
)
File "/usr/local/lib/python3.14/site-packages/datasets/inspect.py", line 291, in get_dataset_config_info
raise SplitsNotFoundError("The split names could not be parsed from the dataset config.") from err
datasets.inspect.SplitsNotFoundError: The split names could not be parsed from the dataset config.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.
Relapse: runs, per-question results and logs
Everything behind Relapse: forget the hazard, keep the rest, and check it in the 4-bit file people download. Cohere's Tiny Aya Global, unlearned with Lexsi SafeTune's RMU at five layer windows (30 runs, plus the runs that repair the one window that relapses), scored in up to seven storage formats.
| Folder | What is in it |
|---|---|
raw/grid_*.json |
One row per model and format: WMDP-bio and MMLU accuracy, probability on the right answer, and accuracy on each probe set (probe:hi, probe:ar, probe:es, probe:sw, probe:en_para). grid_seed1.json is the second training seed; grid_deeper.json is the layer-10-to-12 control in real llama.cpp Q4_0 and Q4_K_M files |
raw/sweep_*.json |
The tuning sweeps: SafeTune's RMU defaults, the retune, the CircuitKIT-layer settings, and NPO. sweep_rmu_controls_*.json are the controls: the default layers at a matched budget, and the same edit at layers 3 to 5, 10 to 12 and 13 to 15 |
raw/sweep_rmu_fix_*.json, raw/grid_deeperfix.json |
The repair: the layer-10-to-12 edit retrained with the model seen as its Q4_0 file (on half the steps, on every step, twice as long, a higher learning rate, a second seed), the ablation that rounds only the edited matrices, and real llama.cpp Q4_0 and Q4_K_M files of the repaired model |
raw/variant_*.json |
The three placements (default, CircuitKIT, deeper) on Tiny Aya Fire and Tiny Aya Earth, settings unchanged from Global |
raw/locate_relapse.json |
Where the relapse comes from: WMDP-bio with only one part of the model rounded to Q4_0 (the layers before the edit, the edited matrices, the rest of the window, the layers after) |
raw/edit_vs_rounding.json |
For every run, the size of the edit against the Q4_0 rounding error in the same matrices |
raw/robustness_*.json |
Prompt attacks (five worked examples, an expert persona) and the fine-tuning attack (25 LoRA steps on 64 WMDP-bio questions, and on 64 unrelated MMLU questions as a control), scored on the 1,204 WMDP-bio questions not used as examples |
raw/items_*.json |
Per question: the four answer probabilities on WMDP-bio and the probability on the right answer for MMLU, keyed model|format|set |
raw/audit_*.json |
AuditKIT paired comparisons: forgetting fp16 against Q4_K_M (and Q4_0 in audit_formats_deeper.json), the general-knowledge price against the original at the same format, and the two headline recipes against each other in audit_recipes.json (delta, 95% bootstrap interval, p-value, flips) |
audit_runs/ |
The full AuditKIT runs (every prediction, log-likelihoods, fingerprint) |
raw/circuitkit_layers.json, circuitkit/ |
CircuitKIT EAP-IG attribution on WMDP-bio: per-layer MLP and attention scores, node scores, faithfulness report |
raw/bench_*.json |
llama.cpp CPU and transformers T4 speed measurements used in the cost table |
logs/ |
Kaggle kernel logs of the unlearning runs, the sweeps, the attacks and the final grids |
Run names map to the write-up's methods: base = original, rmu_default_layers = RMU at SafeTune's default layers 5 to 7,
rmu_circuit_layers = RMU at the CircuitKIT layers 7 to 9, rmu_circuit_layers_qa = the same trained through a 4-bit
quantizer, rmu_L7_strong = RMU at default layers with 2.5x the learning rate, *_seed1 = the second seed, rmu_deeper_* and X10-12_* = the control at layers 10 to 12 that relapses in 4-bit files, X3-5_* and X13-15_* = the other control windows, D5-7_* = the default layers at the CircuitKIT runs' budget, X10-12_*_file* and rmu_deeper_file = the repair runs (_file = model seen as its Q4_0 file on half the steps, _file_p1 = on every step), X10-12_lr4.5e-4_qa = only the edited matrices rounded. results/methods.json in the code repository maps every sweep run name to its settings; the edited weights of every run are in itssaharsh/tiny-aya-global-unlearn-bio-sweep. Formats: fp16, bf16, Q8_0, Q4_0 and nf4 are
exact weight round trips; Q6_K and Q4_K_M are real llama.cpp files, and so is Q4_0 in grid_deeper.json and grid_deeperfix.json.
Rebuild every number on the Space: python scripts/build_results.py --fetch in github.com/itssaharsh/relapse.
- Downloads last month
- 459