2026 MATS: rational agents
Collection
The 2026 MATS rational-agents programme and its sibling MATS-II experiments: curated 2026.RA.* repos plus the raw run archives (Oct 2026). • 55 items • Updated
Dataset Viewer
The dataset viewer is not available for this split.
Cannot load the dataset split (in streaming mode) to extract the first rows.
Error code: StreamingRowsError
Exception: CastError
Message: Couldn't cast
</think>: int64
</tool_call>: int64
</tool_response>: int64
<think>: int64
<tool_call>: int64
<tool_response>: int64
<|box_end|>: int64
<|box_start|>: int64
<|endoftext|>: int64
<|file_sep|>: int64
<|fim_middle|>: int64
<|fim_pad|>: int64
<|fim_prefix|>: int64
<|fim_suffix|>: int64
<|im_end|>: int64
<|im_start|>: int64
<|image_pad|>: int64
<|object_ref_end|>: int64
<|object_ref_start|>: int64
<|quad_end|>: int64
<|quad_start|>: int64
<|repo_name|>: int64
<|video_pad|>: int64
<|vision_end|>: int64
<|vision_pad|>: int64
<|vision_start|>: int64
use_dora: bool
lora_dropout: double
target_modules: list<item: string>
child 0, item: string
lora_alpha: int64
use_bdlora: null
base_model_name_or_path: string
bias: string
init_lora_weights: bool
use_rslora: bool
lora_ga_config: null
peft_version: string
trainable_token_indices: null
alora_invocation_tokens: null
auto_mapping: null
peft_type: string
revision: null
use_qalora: bool
rank_pattern: struct<>
r: int64
corda_config: null
layers_pattern: null
fan_in_fan_out: bool
qalora_group_size: int64
eva_config: null
arrow_config: null
lora_bias: bool
alpha_pattern: struct<>
ensure_weight_tying: bool
layer_replication: null
layers_to_transform: null
target_parameters: null
task_type: string
modules_to_save: null
inference_mode: bool
exclude_modules: null
megatron_config: null
loftq_config: struct<>
megatron_core: string
to
{'alora_invocation_tokens': Value('null'), 'alpha_pattern': {}, 'arrow_config': Value('null'), 'auto_mapping': Value('null'), 'base_model_name_or_path': Value('string'), 'bias': Value('string'), 'corda_config': Value('null'), 'ensure_weight_tying': Value('bool'), 'eva_config': Value('null'), 'exclude_modules': Value('null'), 'fan_in_fan_out': Value('bool'), 'inference_mode': Value('bool'), 'init_lora_weights': Value('bool'), 'layer_replication': Value('null'), 'layers_pattern': Value('null'), 'layers_to_transform': Value('null'), 'loftq_config': {}, 'lora_alpha': Value('int64'), 'lora_bias': Value('bool'), 'lora_dropout': Value('float64'), 'lora_ga_config': Value('null'), 'megatron_config': Value('null'), 'megatron_core': Value('string'), 'modules_to_save': Value('null'), 'peft_type': Value('string'), 'peft_version': Value('string'), 'qalora_group_size': Value('int64'), 'r': Value('int64'), 'rank_pattern': {}, 'revision': Value('null'), 'target_modules': List(Value('string')), 'target_parameters': Value('null'), 'task_type': Value('string'), 'trainable_token_indices': Value('null'), 'use_bdlora': Value('null'), 'use_dora': Value('bool'), 'use_qalora': Value('bool'), 'use_rslora': Value('bool')}
because column names don't match
Traceback: Traceback (most recent call last):
File "/src/services/worker/src/worker/utils.py", line 147, in get_rows_or_raise
return get_rows(
dataset=dataset,
...<4 lines>...
column_names=column_names,
)
File "/src/libs/libcommon/src/libcommon/utils.py", line 272, in decorator
return func(*args, **kwargs)
File "/src/services/worker/src/worker/utils.py", line 127, in get_rows
rows_plus_one = list(itertools.islice(safe_iter(ds, dataset=dataset), rows_max_number + 1))
File "/src/services/worker/src/worker/utils.py", line 483, in safe_iter
yield from ds.decode(False) if ds.features else ds
File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 2840, in __iter__
for key, example in ex_iterable:
^^^^^^^^^^^
File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 2373, in __iter__
for key, pa_table in self._iter_arrow():
~~~~~~~~~~~~~~~~^^
File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 2398, in _iter_arrow
for key, pa_table in self.ex_iterable._iter_arrow():
~~~~~~~~~~~~~~~~~~~~~~~~~~~~^^
File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 536, in _iter_arrow
for key, pa_table in iterator:
^^^^^^^^
File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 419, 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 343, 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 2378, in table_cast
return cast_table_to_schema(table, schema)
File "/usr/local/lib/python3.14/site-packages/datasets/table.py", line 2306, in cast_table_to_schema
raise CastError(
...<3 lines>...
)
datasets.table.CastError: Couldn't cast
</think>: int64
</tool_call>: int64
</tool_response>: int64
<think>: int64
<tool_call>: int64
<tool_response>: int64
<|box_end|>: int64
<|box_start|>: int64
<|endoftext|>: int64
<|file_sep|>: int64
<|fim_middle|>: int64
<|fim_pad|>: int64
<|fim_prefix|>: int64
<|fim_suffix|>: int64
<|im_end|>: int64
<|im_start|>: int64
<|image_pad|>: int64
<|object_ref_end|>: int64
<|object_ref_start|>: int64
<|quad_end|>: int64
<|quad_start|>: int64
<|repo_name|>: int64
<|video_pad|>: int64
<|vision_end|>: int64
<|vision_pad|>: int64
<|vision_start|>: int64
use_dora: bool
lora_dropout: double
target_modules: list<item: string>
child 0, item: string
lora_alpha: int64
use_bdlora: null
base_model_name_or_path: string
bias: string
init_lora_weights: bool
use_rslora: bool
lora_ga_config: null
peft_version: string
trainable_token_indices: null
alora_invocation_tokens: null
auto_mapping: null
peft_type: string
revision: null
use_qalora: bool
rank_pattern: struct<>
r: int64
corda_config: null
layers_pattern: null
fan_in_fan_out: bool
qalora_group_size: int64
eva_config: null
arrow_config: null
lora_bias: bool
alpha_pattern: struct<>
ensure_weight_tying: bool
layer_replication: null
layers_to_transform: null
target_parameters: null
task_type: string
modules_to_save: null
inference_mode: bool
exclude_modules: null
megatron_config: null
loftq_config: struct<>
megatron_core: string
to
{'alora_invocation_tokens': Value('null'), 'alpha_pattern': {}, 'arrow_config': Value('null'), 'auto_mapping': Value('null'), 'base_model_name_or_path': Value('string'), 'bias': Value('string'), 'corda_config': Value('null'), 'ensure_weight_tying': Value('bool'), 'eva_config': Value('null'), 'exclude_modules': Value('null'), 'fan_in_fan_out': Value('bool'), 'inference_mode': Value('bool'), 'init_lora_weights': Value('bool'), 'layer_replication': Value('null'), 'layers_pattern': Value('null'), 'layers_to_transform': Value('null'), 'loftq_config': {}, 'lora_alpha': Value('int64'), 'lora_bias': Value('bool'), 'lora_dropout': Value('float64'), 'lora_ga_config': Value('null'), 'megatron_config': Value('null'), 'megatron_core': Value('string'), 'modules_to_save': Value('null'), 'peft_type': Value('string'), 'peft_version': Value('string'), 'qalora_group_size': Value('int64'), 'r': Value('int64'), 'rank_pattern': {}, 'revision': Value('null'), 'target_modules': List(Value('string')), 'target_parameters': Value('null'), 'task_type': Value('string'), 'trainable_token_indices': Value('null'), 'use_bdlora': Value('null'), 'use_dora': Value('bool'), 'use_qalora': Value('bool'), 'use_rslora': Value('bool')}
because column names don't matchNeed help to make the dataset viewer work? Make sure to review how to configure the dataset viewer, and open a discussion for direct support.
2026.RA.TimeCost-GRPO — pricing the clock in a negotiation GRPO reward
Campaign artifacts for time-cost GRPO (Qwen3-8B LoRA negotiation policy). Every reward this program has
trained on is blind to the clock: GameSpec.discount = 0.95 is declared in the spec and applied to nothing, so
a deal struck in round 1 and the same deal struck at the buzzer score identically and every no-deal pays the
same constant — grinding to the deadline is exactly free, which is the Sandholm-Vulkan brinkmanship baseline.
This campaign adds R_i = g(z_i) - c * tau, tau = (r_end - 1) / R the elapsed fraction of the round
deadline, and asks whether that removes the deadline-attrition attractor and whether it composes with
branch-backtracking (2026.RA.Branch-Backtrack-GRPO).
c = 0.5 g-units over the whole clock and horizon no-deal billing were frozen on CPU before any GPU
(phase0/): half the analytic ordering ceiling, and horizon because among stored no-deal episodes ending
early is equivalent to somebody walked (4 of 452), so billing a no-deal for the clock it consumed would pay a
bonus for quitting sooner. All gates were preregistered in the source repo before training.
Two arms, identical but for the branch schedule: tc25 (25 standard waves) and tcbranch25
(--branch-every 2 --branch-groups 6 --branch-last-rounds 4, whole-game nodes).
Verdict: the preregistered POSITIVE gate FAILS, and the endpoint moved the wrong way
Adjudicated at rung 15 on the held-out primary bank (24 game clusters, 48 instances x 2 protocol arms x 10
seeds), paired on (instance, seed, arm) against a hardware-matched base re-run on the same a6000 fleet,
cluster-bootstrapped over games. Every number below is read from analysis/verdict_rung15.json at build time.
| arm (rung 15) | deal (G1) | at-cap (G2) | walk (G3) | rounds-to-close |
|---|---|---|---|---|
| tc25_r15 | -0.439 [-0.488, -0.391] | +0.512 [+0.458, +0.561] | +0.149 [+0.113, +0.189] | +1.623 [+1.386, +1.855] |
| tcbranch25_r15 | -0.267 [-0.309, -0.224] | +0.463 [+0.420, +0.505] | +0.054 [+0.029, +0.081] | +1.476 [+1.311, +1.635] |
Closure collapses, attrition explodes, and deals close LATER — under a reward that subtracts 0.5 * tau
for elapsed rounds. That last column is the campaign's own question answering itself backwards: a policy
charged for time spent about a round and a half more of it. The training-side manipulation check (H1,
time/elapsed_fraction_deals) rises rather than falls in both arms, so this campaign cannot say why — a
time term that was dominated and one that never reached the policy are not distinguished by anything measured
here. Pricing the clock is not what was missing. Three distinct reward interventions have now collapsed
closure at this rung on this bank (lam=0, branch-backtracking, and this one), each failing its own gate
against its own matched base — the earlier numbers are historical context on a different hardware era and
are deliberately NOT paired against. What this campaign adds on its own terms: the branch dose reached 17.4%
of gradient mass, 1.7x the ~10% an 80 GB card's OOM left the branch campaign with, and closure still
collapsed (branch arm G1 -0.267 [-0.309, -0.224], FAIL). "Not enough branch gradient" is therefore no longer
available as an explanation for branching failing to rescue closure.
A composition claim was available and is NOT made. The branch-vs-no-branch contrast fires its
preregistered test (deal +0.194 [+0.141, +0.252], CI excluding 0), but the two arms differ ~4x in eval-time
format collapse — policy-turn parse failure 15.24% vs 3.89%, episodes carrying a void turn 96.2% vs 64.4%,
against a 0.85% / 10.2% base — which produces a closure gap in the same direction with no negotiation
mechanism involved. Restricted to void-free episodes the two arms are level. What is supported: at identical
time-cost dose, branch training left the policy degrading its output format ~4x more slowly. See
analysis/void_confound_rung15.json; the probe is post-treatment selection and is not a corrected estimate.
Full per-contrast detail:
{
"tc25_r15": {
"deal": {
"estimate": -0.4392361111111111,
"ci_low": -0.48789062499999997,
"ci_high": -0.39058159722222247,
"n_pairs": 576,
"n_clusters": 48,
"pairing": "instance+seed+arm"
},
"at_cap": {
"estimate": 0.5121527777777778,
"ci_low": 0.4583333333333333,
"ci_high": 0.5607638888888888,
"n_pairs": 576,
"n_clusters": 48,
"pairing": "instance+seed+arm"
},
"walk": {
"estimate": 0.14930555555555555,
"ci_low": 0.11284722222222222,
"ci_high": 0.1892361111111111,
"n_pairs": 576,
"n_clusters": 48,
"pairing": "instance+seed+arm"
},
"below_threshold": {
"estimate": -0.20833333333333334,
"ci_low": -0.2552083333333333,
"ci_high": -0.16493055555555555,
"n_pairs": 576,
"n_clusters": 48,
"pairing": "instance+seed+arm",
"VOID": true
},
"tau_deal": {
"estimate": 0.4057203389830508,
"ci_low": 0.34639613434287303,
"ci_high": 0.46365771990905214,
"n_pairs": 236,
"n_clusters": 48,
"pairing": "instance+seed+arm",
"VOID": true
},
"round_close": {
"estimate": 1.6228813559322033,
"ci_low": 1.3855845373714921,
"ci_high": 1.8546308796362085,
"n_pairs": 236,
"n_clusters": 48,
"pairing": "instance+seed+arm",
"VOID": true
},
"at_round_cap": {
"estimate": 0.4722222222222222,
"ci_low": 0.4236111111111111,
"ci_high": 0.5208333333333334,
"n_pairs": 576,
"n_clusters": 48,
"pairing": "instance+seed+arm"
}
},
"tcbranch25_r15": {
"deal": {
"estimate": -0.26666666666666666,
"ci_low": -0.309375,
"ci_high": -0.2239322916666668,
"n_pairs": 960,
"n_clusters": 48,
"pairing": "instance+seed+arm"
},
"at_cap": {
"estimate": 0.4625,
"ci_low": 0.4197916666666667,
"ci_high": 0.5052343749999999,
"n_pairs": 960,
"n_clusters": 48,
"pairing": "instance+seed+arm"
},
"walk": {
"estimate": 0.05416666666666667,
"ci_low": 0.029140625000000003,
"ci_high": 0.08125,
"n_pairs": 960,
"n_clusters": 48,
"pairing": "instance+seed+arm"
},
"below_threshold": {
"estimate": -0.190625,
"ci_low": -0.23645833333333333,
"ci_high": -0.14583333333333334,
"n_pairs": 960,
"n_clusters": 48,
"pairing": "instance+seed+arm",
"VOID": true
},
"tau_deal": {
"estimate": 0.36906934306569344,
"ci_low": 0.32782540631961055,
"ci_high": 0.40873503773093933,
"n_pairs": 548,
"n_clusters": 48,
"pairing": "instance+seed+arm",
"VOID": true
},
"round_close": {
"estimate": 1.4762773722627738,
"ci_low": 1.3113016252784422,
"ci_high": 1.6349401509237573,
"n_pairs": 548,
"n_clusters": 48,
"pairing": "instance+seed+arm",
"VOID": true
},
"at_round_cap": {
"estimate": 0.428125,
"ci_low": 0.38333333333333336,
"ci_high": 0.478125,
"n_pairs": 960,
"n_clusters": 48,
"pairing": "instance+seed+arm"
}
},
"tcbranch25_r15_seedmatched": {
"deal": {
"estimate": -0.24479166666666666,
"ci_low": -0.2951388888888889,
"ci_high": -0.19618055555555555,
"n_pairs": 576,
"n_clusters": 48,
"pairing": "instance+seed+arm"
},
"at_cap": {
"estimate": 0.4600694444444444,
"ci_low": 0.4079861111111111,
"ci_high": 0.5138888888888888,
"n_pairs": 576,
"n_clusters": 48,
"pairing": "instance+seed+arm"
},
"walk": {
"estimate": 0.0642361111111111,
"ci_low": 0.03125,
"ci_high": 0.09895833333333333,
"n_pairs": 576,
"n_clusters": 48,
"pairing": "instance+seed+arm"
},
"below_threshold": {
"estimate": -0.19270833333333334,
"ci_low": -0.2465277777777778,
"ci_high": -0.1423611111111111,
"n_pairs": 576,
"n_clusters": 48,
"pairing": "instance+seed+arm",
"VOID": true
},
"tau_deal": {
"estimate": 0.35058309037900875,
"ci_low": 0.2984730021507896,
"ci_high": 0.3994401070380485,
"n_pairs": 343,
"n_clusters": 48,
"pairing": "instance+seed+arm",
"VOID": true
},
"round_close": {
"estimate": 1.402332361516035,
"ci_low": 1.1938920086031584,
"ci_high": 1.597760428152194,
"n_pairs": 343,
"n_clusters": 48,
"pairing": "instance+seed+arm",
"VOID": true
},
"at_round_cap": {
"estimate": 0.4201388888888889,
"ci_low": 0.36974826388888893,
"ci_high": 0.4722222222222222,
"n_pairs": 576,
"n_clusters": 48,
"pairing": "instance+seed+arm"
}
}
}
Cell levels:
{
"baseline": {
"deal": 0.8989583333333333,
"nnw": 0.6812588541666666,
"below_threshold": 0.31979166666666664,
"at_cap": 0.375,
"among_ir": 0.3385416666666667,
"nnw_among_ir": 0.8430513846153848,
"gini": 0.33419604166666667,
"worst_share": -0.025177469559158956,
"max_share": 0.36864315588544033,
"walk": 0.042708333333333334,
"expiry": 0.058333333333333334,
"gen_failures": 0.0,
"tau": 0.6231770833333333,
"tau_deal": 0.5848783314020858,
"round_close": 3.339513325608343,
"at_round_cap": 0.5
},
"tc25_r15": {
"deal": 0.4600694444444444,
"nnw": 0.36554184027777775,
"below_threshold": 0.109375,
"at_cap": 0.8836805555555556,
"among_ir": 0.1701388888888889,
"nnw_among_ir": 0.8471265306122447,
"gini": 0.17143802083333334,
"worst_share": -0.002360932789129282,
"max_share": 0.3428136908042625,
"walk": 0.1892361111111111,
"expiry": 0.3506944444444444,
"gen_failures": 0.0,
"tau": 0.9869791666666666,
"tau_deal": 0.9716981132075472,
"round_close": 4.886792452830188,
"at_round_cap": 0.9756944444444444
},
"tcbranch25_r15": {
"deal": 0.6322916666666667,
"nnw": 0.5291611458333333,
"below_threshold": 0.12916666666666668,
"at_cap": 0.8375,
"among_ir": 0.2791666666666667,
"nnw_among_ir": 0.8853264925373135,
"gini": 0.22834364583333333,
"worst_share": -0.0007285087228761062,
"max_share": 0.3329143389863073,
"walk": 0.096875,
"expiry": 0.2708333333333333,
"gen_failures": 0.0,
"tau": 0.95859375,
"tau_deal": 0.9365733113673805,
"round_close": 4.746293245469523,
"at_round_cap": 0.928125
},
"tcbranch25_r15_seedmatched": {
"deal": 0.6545138888888888,
"nnw": 0.5495758680555556,
"below_threshold": 0.125,
"at_cap": 0.8315972222222222,
"among_ir": 0.28125,
"nnw_among_ir": 0.8777598765432099,
"gini": 0.24071180555555557,
"worst_share": -0.0015618505822816834,
"max_share": 0.3337404755767845,
"walk": 0.10416666666666667,
"expiry": 0.24131944444444445,
"gen_failures": 0.0,
"tau": 0.9578993055555556,
"tau_deal": 0.9363395225464191,
"round_close": 4.745358090185676,
"at_round_cap": 0.9236111111111112
}
}
Contents: phase0/ (the calibration that froze c, plus per-row rescores of 7,288 focal-seat rows),
pool/ (618 replay-verified whole-game branch nodes + the soundness gate), training/ (both arms: telemetry,
wave transcripts, LoRA checkpoints), eval/ (one jsonl.gz per COMPLETE eval shard — partial shards excluded,
since censoring would drop exactly the longest episodes), analysis/ (the paired contrasts and gate verdicts).
Episode schema: interlens Episode.to_json().
Build counts (negative = shard excluded as incomplete):
{
"phase0/rows": 7288,
"analysis/files": 8,
"tc25/step0005": 96,
"tc25/step0010": 96,
"tc25/step0015": 96,
"tcbranch25/step0005": 96,
"tcbranch25/step0010": 48,
"tcbranch25/step0015": 96,
"tcbranch25/step0020": 48,
"tcbranch25/step0025": 96,
"eval/base_primary_s0": 96,
"eval/base_primary_s1": 96,
"eval/base_primary_s2": 96,
"eval/base_primary_s3": 96,
"eval/base_primary_s4": 96,
"eval/base_primary_s5": 96,
"eval/base_primary_s6": 96,
"eval/base_primary_s7": 96,
"eval/base_primary_s8": 96,
"eval/base_primary_s9": 96,
"eval/tc25_r15_primary_s0": 96,
"eval/tc25_r15_primary_s1": 96,
"eval/tc25_r15_primary_s2": 96,
"eval/tc25_r15_primary_s3": 96,
"eval/tc25_r15_primary_s4": 96,
"eval/tc25_r15_primary_s5": 96,
"eval/tc25_r15_primary_s6": -34,
"eval/tc25_r15_primary_s7": -30,
"eval/tc25_r15_primary_s8": -4,
"eval/tc25_r15_primary_s9": -30,
"eval/tc25_r5_primary_s0": 96,
"eval/tc25_r5_primary_s1": 96,
"eval/tc25_r5_primary_s2": 96,
"eval/tc25_r5_primary_s3": 96,
"eval/tc25_r5_primary_s4": 96,
"eval/tc25_r5_primary_s5": 96,
"eval/tc25_r5_primary_s6": 96,
"eval/tc25_r5_primary_s7": 96,
"eval/tc25_r5_primary_s8": 96,
"eval/tc25_r5_primary_s9": 96,
"eval/tcbranch25_r15_primary_s0": 96,
"eval/tcbranch25_r15_primary_s1": 96,
"eval/tcbranch25_r15_primary_s2": 96,
"eval/tcbranch25_r15_primary_s3": 96,
"eval/tcbranch25_r15_primary_s4": 96,
"eval/tcbranch25_r15_primary_s5": 96,
"eval/tcbranch25_r15_primary_s6": 96,
"eval/tcbranch25_r15_primary_s7": 96,
"eval/tcbranch25_r15_primary_s8": 96,
"eval/tcbranch25_r15_primary_s9": 96,
"eval/tcbranch25_r5_primary_s0": 96,
"eval/tcbranch25_r5_primary_s1": 96,
"eval/tcbranch25_r5_primary_s2": 96,
"eval/tcbranch25_r5_primary_s3": 96,
"eval/tcbranch25_r5_primary_s4": 96,
"eval/tcbranch25_r5_primary_s5": 96,
"eval/tcbranch25_r5_primary_s6": 96,
"eval/tcbranch25_r5_primary_s7": 96,
"eval/tcbranch25_r5_primary_s8": 96,
"eval/tcbranch25_r5_primary_s9": 96
}
- Downloads last month
- 53