Datasets:

JAEMP
/

repro-knowing-when-to-quit-bundle

model stringclasses 5 values	abstention_ratio float64 0.1 0.9	accuracy float64 0.04 1	acc_ci float64 0 0.15	reward float64 0.16 1	rew_ci float64 0 0.51	saved_tokens float64 0 0.93	saved_tokens_ci float64 0 0.04
full	0.1	0.177881	0.001143	0.172135	0.00883	0.061566	0.004217
full	0.2	0.194444	0.002272	0.180833	0.023552	0.125139	0.003505
full	0.3	0.213228	0.002219	0.238284	0.146682	0.189767	0.002098
full	0.4	0.239506	0.005929	0.181226	0.031887	0.262388	0.004477
full	0.5	0.270185	0.007469	0.196811	0.043926	0.343622	0.008171
full	0.6	0.31412	0.006313	0.312597	0.299236	0.434789	0.008875
full	0.7	0.379938	0.00496	0.294873	0.201416	0.533211	0.013253
full	0.8	0.482599	0.012884	0.220618	0.037959	0.645671	0.018527
full	0.9	0.642791	0.013788	0.517866	0.506446	0.786508	0.015665
no_abstention	0.1	0.164815	0	0.164815	0	0	0
no_abstention	0.2	0.164815	0	0.164815	0	0	0
no_abstention	0.3	0.164815	0	0.164815	0	0	0
no_abstention	0.4	0.164815	0	0.164815	0	0	0
no_abstention	0.5	0.164815	0	0.164815	0	0	0
no_abstention	0.6	0.164815	0	0.164815	0	0	0
no_abstention	0.7	0.164815	0	0.164815	0	0	0
no_abstention	0.8	0.164815	0	0.164815	0	0	0
no_abstention	0.9	0.164815	0	0.164815	0	0	0
full	0.1	0.468519	0.001653	0.439915	0.012727	0.056884	0.002982
full	0.2	0.509144	0.005868	0.478495	0.093601	0.111383	0.002596
full	0.3	0.551455	0.007669	0.571852	0.145733	0.170405	0.007288
full	0.4	0.59784	0.010932	0.659656	0.170853	0.240094	0.007316
full	0.5	0.652222	0.007738	0.726111	0.170276	0.314957	0.007951
full	0.6	0.71713	0.005241	0.67621	0.253567	0.401584	0.01221
full	0.7	0.79537	0.017852	0.524389	0.075098	0.502063	0.010535
full	0.8	0.861717	0.015594	0.758568	0.275538	0.620339	0.012706
full	0.9	0.906977	0.019585	0.804698	0.241633	0.753308	0.010862
no_abstention	0.1	0.430093	0	0.430093	0	0	0
no_abstention	0.2	0.430093	0	0.430093	0	0	0
no_abstention	0.3	0.430093	0	0.430093	0	0	0
no_abstention	0.4	0.430093	0	0.430093	0	0	0
no_abstention	0.5	0.430093	0	0.430093	0	0	0
no_abstention	0.6	0.430093	0	0.430093	0	0	0
no_abstention	0.7	0.430093	0	0.430093	0	0	0
no_abstention	0.8	0.430093	0	0.430093	0	0	0
no_abstention	0.9	0.430093	0	0.430093	0	0	0
baseline	0.1	0.179835	0.001895	0.168281	0.007012	0.126042	0.004293
baseline	0.2	0.198611	0.001181	0.195629	0.048837	0.242256	0.006237
baseline	0.3	0.219312	0.003561	0.230931	0.061808	0.357511	0.014873
baseline	0.4	0.240741	0.018664	0.242845	0.113455	0.469673	0.02961
baseline	0.5	0.256481	0.038878	0.321044	0.224584	0.568924	0.028684
baseline	0.6	0.274769	0.048731	0.375279	0.155293	0.664733	0.026899
baseline	0.7	0.29537	0.075701	0.685889	0.283659	0.763565	0.022919
baseline	0.8	0.305336	0.083759	0.46728	0.304478	0.849736	0.01522
baseline	0.9	0.341395	0.090727	0.547896	0.283755	0.931013	0.006096
full	0.1	0.180967	0.00159	0.16936	0.003354	0.079248	0.004959
full	0.2	0.198958	0.001558	0.188852	0.048745	0.165203	0.006737
full	0.3	0.222487	0.001374	0.235286	0.154432	0.251353	0.003558
full	0.4	0.255864	0.001987	0.217918	0.097816	0.349034	0.008762
full	0.5	0.300741	0.001259	0.286214	0.259667	0.454441	0.012909
full	0.6	0.361343	0.003116	0.194535	0.029974	0.571859	0.013039
full	0.7	0.437654	0.007617	0.19753	0.015144	0.686498	0.011054
full	0.8	0.538283	0.025115	0.412033	0.374972	0.778249	0.011097
full	0.9	0.645581	0.049411	0.532987	0.346936	0.877265	0.009202
lora_abstention	0.1	0.159465	0	0.17182	0	0.098078	0
lora_abstention	0.2	0.157407	0	0.17037	0	0.193574	0
lora_abstention	0.3	0.158069	0	0.410648	0	0.293136	0
lora_abstention	0.4	0.158179	0	0.165439	0	0.390704	0
lora_abstention	0.5	0.153704	0	0.276852	0	0.49512	0
lora_abstention	0.6	0.172454	0	0.668981	0	0.597477	0
lora_abstention	0.7	0.171296	0	0.751389	0	0.695777	0
lora_abstention	0.8	0.160093	0	0.192952	0	0.798541	0
lora_abstention	0.9	0.172093	0	0.267209	0	0.901678	0
no_abstention	0.1	0.164815	0	0.164815	0	0	0
no_abstention	0.2	0.164815	0	0.164815	0	0	0
no_abstention	0.3	0.164815	0	0.164815	0	0	0
no_abstention	0.4	0.164815	0	0.164815	0	0	0
no_abstention	0.5	0.164815	0	0.164815	0	0	0
no_abstention	0.6	0.164815	0	0.164815	0	0	0
no_abstention	0.7	0.164815	0	0.164815	0	0	0
no_abstention	0.8	0.164815	0	0.164815	0	0	0
no_abstention	0.9	0.164815	0	0.164815	0	0	0
self_assessment	0.1	0.151235	0	0.169877	0	0.077888	0
self_assessment	0.2	0.137731	0	0.210185	0	0.164558	0
self_assessment	0.3	0.132937	0	0.183056	0	0.261941	0
self_assessment	0.4	0.131944	0	0.479167	0	0.371678	0
self_assessment	0.5	0.124074	0	0.17698	0	0.466346	0
self_assessment	0.6	0.086806	0	0.217956	0	0.54508	0
self_assessment	0.7	0.07716	0	0.168603	0	0.647703	0
self_assessment	0.8	0.044084	0	0.165942	0	0.76345	0
self_assessment	0.9	0.07907	0	0.232907	0	0.882587	0
baseline	0.1	0.43786	0.005255	0.475471	0.023589	0.100142	0.006233
baseline	0.2	0.444213	0.015129	0.540634	0.045531	0.19976	0.006411
baseline	0.3	0.451852	0.032034	0.483933	0.096568	0.299715	0.010931
baseline	0.4	0.464352	0.0474	0.598167	0.128692	0.404621	0.019103
baseline	0.5	0.473148	0.06035	0.645398	0.163376	0.50867	0.020298
baseline	0.6	0.480556	0.062356	0.601839	0.188299	0.610171	0.018849
baseline	0.7	0.504012	0.070905	0.690383	0.26575	0.71214	0.011623
baseline	0.8	0.566589	0.082406	0.686219	0.265929	0.817244	0.009101
baseline	0.9	0.694884	0.150853	0.657577	0.22313	0.914369	0.004664
full	0.1	0.463992	0.001498	0.445075	0.017553	0.065203	0.007108
full	0.2	0.510417	0.001832	0.448473	0.013514	0.140188	0.005077
full	0.3	0.561111	0.004879	0.506259	0.133444	0.22827	0.005166
full	0.4	0.617438	0.007464	0.53671	0.12124	0.329419	0.011199
full	0.5	0.687778	0.005828	0.559107	0.211591	0.441767	0.007089
full	0.6	0.774074	0.011011	0.570948	0.144508	0.560756	0.01332
full	0.7	0.861728	0.008197	0.601267	0.250669	0.681013	0.017215
full	0.8	0.912761	0.005974	0.638816	0.241434	0.783502	0.009831
full	0.9	0.913488	0.005165	0.991349	0.000517	0.871799	0.005381
lora_abstention	0.1	0.442901	0.011949	0.456014	0.034876	0.097543	0.007252

End of preview. Expand in Data Studio

YAML Metadata Warning:empty or missing yaml metadata in repo card

Check out the documentation for more information.

Reproduction bundle

This bundle contains the executable local-GPU scaled reproduction, the official metric helper, generated metrics, and the poster artifact.

Sources

Paper: https://huggingface.co/papers/2604.18419
Official code: https://github.com/HenDav/mid-gen-abs
Official trajectory dataset: https://huggingface.co/datasets/HenDav/mid-gen-abs-data
Direct model: https://huggingface.co/microsoft/Phi-3-mini-4k-instruct
Transfer dataset: https://huggingface.co/datasets/Hothan/OlympiadBench

Rerun

uv sync
uv run python direct_model_reproduction.py

For the artifact re-aggregation, download the official trajectory files into traj_csvs/ using download_data.py, then run:

uv run python evaluate_claims.py

The direct model run is intentionally scaled: GSM8K train/eval 64/32 and a 16-example English text-only OlympiadBench transfer slice. It does not claim the paper-scale 64%/91% headline reproduction.

Downloads last month: 71

Paper for JAEMP/repro-knowing-when-to-quit-bundle

Knowing When to Quit: A Principled Framework for Dynamic Abstention in LLM Reasoning

Paper • 2604.18419 • Published Jun 12 • 3