dynamic-qwen-vpr-gamma0.5

fredericowieser/dynamic-qwen-vpr-gamma0.5

dynamic-qwen-vpr-gamma0.5 is a machine learning model from fredericowieser. Use it for the machine learning task on the model card, and read the license before you ship it in a product. The card lists the license as apache-2.0.

This model is a dynamically-computed version of outputs/PRETRAIN-TEST-qwen2.5-0.5B-vpr-pretrainmix-2025-08-2813-29-10-gamma=0.5/finalmodel, fine-tuned using the VPR architecture.

Downloads · 30 days

17% of all-time downloads

All-time downloads

Public

Parameters

496M

2 GB on disk

Likes

Public

Hugging Face

Repo makeup

Click a slice to open those files.

.safetensors2 GB · 99%

At a glance

License: apache-2.0
Model type: dynamic_qwen
Access: Public
Created: Aug 29, 2025
Updated: Aug 29, 2025
SHA: 03d4d5c7

Model card

From the Hugging Face model README

Model Card for dynamic-qwen-vpr-gamma0.5

This model is a dynamically-computed version of outputs/PRETRAIN-TEST-qwen2.5-0.5B-vpr-pretrain_mix-2025-08-28_13-29-10-gamma=0.5/final_model, fine-tuned using the VPR architecture.

Dynamic Architecture: VPR
Capacity Gamma (γ): 0.5

The VPR architecture enables the model to conditionally skip parts of its computation, aiming for improved efficiency. The capacity_gamma parameter controls the portion of tokens processed by the dynamic components.

How to Use

This model requires trust_remote_code=True to load the custom architecture.

from transformers import AutoModelForCausalLM, AutoTokenizer

# It is recommended to load in bfloat16 for efficiency
model = AutoModelForCausalLM.from_pretrained(
    "fredericowieser/dynamic-qwen-vpr-gamma0.5",
    trust_remote_code=True,
    torch_dtype="auto",
    device_map="auto"
)
tokenizer = AutoTokenizer.from_pretrained("fredericowieser/dynamic-qwen-vpr-gamma0.5")

# Example usage
prompt = "The capital of the United Kingdom is"
inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
outputs = model.generate(**inputs, max_new_tokens=10)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

Evaluation

Results on standard benchmarks:

Task	Metric	Value
arc_challenge	acc,none	0.2244
arc_challenge	acc_norm,none	0.2585
arc_challenge	acc_norm_stderr,none	0.0128
arc_challenge	acc_stderr,none	0.0122
hellaswag	acc,none	0.3020
hellaswag	acc_norm,none	0.3328
hellaswag	acc_norm_stderr,none	0.0047
hellaswag	acc_stderr,none	0.0046
mmlu	acc,none	0.2441
mmlu	acc_stderr,none	0.0036
mmlu_abstract_algebra	acc,none	0.2500
mmlu_abstract_algebra	acc_stderr,none	0.0435
mmlu_anatomy	acc,none	0.3111
mmlu_anatomy	acc_stderr,none	0.0400
mmlu_astronomy	acc,none	0.2303
mmlu_astronomy	acc_stderr,none	0.0343
mmlu_business_ethics	acc,none	0.2200
mmlu_business_ethics	acc_stderr,none	0.0416
mmlu_clinical_knowledge	acc,none	0.2113
mmlu_clinical_knowledge	acc_stderr,none	0.0251
mmlu_college_biology	acc,none	0.2569
mmlu_college_biology	acc_stderr,none	0.0365
mmlu_college_chemistry	acc,none	0.2500
mmlu_college_chemistry	acc_stderr,none	0.0435
mmlu_college_computer_science	acc,none	0.2300
mmlu_college_computer_science	acc_stderr,none	0.0423
mmlu_college_mathematics	acc,none	0.2800
mmlu_college_mathematics	acc_stderr,none	0.0451
mmlu_college_medicine	acc,none	0.2717
mmlu_college_medicine	acc_stderr,none	0.0339
mmlu_college_physics	acc,none	0.2745
mmlu_college_physics	acc_stderr,none	0.0444
mmlu_computer_security	acc,none	0.2300
mmlu_computer_security	acc_stderr,none	0.0423
mmlu_conceptual_physics	acc,none	0.2468
mmlu_conceptual_physics	acc_stderr,none	0.0282
mmlu_econometrics	acc,none	0.2193
mmlu_econometrics	acc_stderr,none	0.0389
mmlu_electrical_engineering	acc,none	0.2345
mmlu_electrical_engineering	acc_stderr,none	0.0353
mmlu_elementary_mathematics	acc,none	0.2540
mmlu_elementary_mathematics	acc_stderr,none	0.0224
mmlu_formal_logic	acc,none	0.1905
mmlu_formal_logic	acc_stderr,none	0.0351
mmlu_global_facts	acc,none	0.2100
mmlu_global_facts	acc_stderr,none	0.0409
mmlu_high_school_biology	acc,none	0.2387
mmlu_high_school_biology	acc_stderr,none	0.0243
mmlu_high_school_chemistry	acc,none	0.2315
mmlu_high_school_chemistry	acc_stderr,none	0.0297
mmlu_high_school_computer_science	acc,none	0.1700
mmlu_high_school_computer_science	acc_stderr,none	0.0378
mmlu_high_school_european_history	acc,none	0.2121
mmlu_high_school_european_history	acc_stderr,none	0.0319
mmlu_high_school_geography	acc,none	0.1869
mmlu_high_school_geography	acc_stderr,none	0.0278
mmlu_high_school_government_and_politics	acc,none	0.2798
mmlu_high_school_government_and_politics	acc_stderr,none	0.0324
mmlu_high_school_macroeconomics	acc,none	0.2282
mmlu_high_school_macroeconomics	acc_stderr,none	0.0213
mmlu_high_school_mathematics	acc,none	0.2333
mmlu_high_school_mathematics	acc_stderr,none	0.0258
mmlu_high_school_microeconomics	acc,none	0.2479
mmlu_high_school_microeconomics	acc_stderr,none	0.0280
mmlu_high_school_physics	acc,none	0.2450
mmlu_high_school_physics	acc_stderr,none	0.0351
mmlu_high_school_psychology	acc,none	0.1963
mmlu_high_school_psychology	acc_stderr,none	0.0170
mmlu_high_school_statistics	acc,none	0.2361
mmlu_high_school_statistics	acc_stderr,none	0.0290
mmlu_high_school_us_history	acc,none	0.2451
mmlu_high_school_us_history	acc_stderr,none	0.0302
mmlu_high_school_world_history	acc,none	0.2658
mmlu_high_school_world_history	acc_stderr,none	0.0288
mmlu_human_aging	acc,none	0.3318
mmlu_human_aging	acc_stderr,none	0.0316
mmlu_human_sexuality	acc,none	0.2366
mmlu_human_sexuality	acc_stderr,none	0.0373
mmlu_humanities	acc,none	0.2440
mmlu_humanities	acc_stderr,none	0.0063
mmlu_international_law	acc,none	0.2562
mmlu_international_law	acc_stderr,none	0.0398
mmlu_jurisprudence	acc,none	0.1852
mmlu_jurisprudence	acc_stderr,none	0.0376
mmlu_logical_fallacies	acc,none	0.3067
mmlu_logical_fallacies	acc_stderr,none	0.0362
mmlu_machine_learning	acc,none	0.2946
mmlu_machine_learning	acc_stderr,none	0.0433
mmlu_management	acc,none	0.2136
mmlu_management	acc_stderr,none	0.0406
mmlu_marketing	acc,none	0.2479
mmlu_marketing	acc_stderr,none	0.0283
mmlu_medical_genetics	acc,none	0.3100
mmlu_medical_genetics	acc_stderr,none	0.0465
mmlu_miscellaneous	acc,none	0.2708
mmlu_miscellaneous	acc_stderr,none	0.0159
mmlu_moral_disputes	acc,none	0.2457
mmlu_moral_disputes	acc_stderr,none	0.0232
mmlu_moral_scenarios	acc,none	0.2212
mmlu_moral_scenarios	acc_stderr,none	0.0139
mmlu_nutrition	acc,none	0.2451
mmlu_nutrition	acc_stderr,none	0.0246
mmlu_other	acc,none	0.2546
mmlu_other	acc_stderr,none	0.0078
mmlu_philosophy	acc,none	0.2797
mmlu_philosophy	acc_stderr,none	0.0255
mmlu_prehistory	acc,none	0.2284
mmlu_prehistory	acc_stderr,none	0.0234
mmlu_professional_accounting	acc,none	0.2305
mmlu_professional_accounting	acc_stderr,none	0.0251
mmlu_professional_law	acc,none	0.2458
mmlu_professional_law	acc_stderr,none	0.0110
mmlu_professional_medicine	acc,none	0.2279
mmlu_professional_medicine	acc_stderr,none	0.0255
mmlu_professional_psychology	acc,none	0.2631
mmlu_professional_psychology	acc_stderr,none	0.0178
mmlu_public_relations	acc,none	0.2000
mmlu_public_relations	acc_stderr,none	0.0383
mmlu_security_studies	acc,none	0.2163
mmlu_security_studies	acc_stderr,none	0.0264
mmlu_social_sciences	acc,none	0.2314
mmlu_social_sciences	acc_stderr,none	0.0076
mmlu_sociology	acc,none	0.2239
mmlu_sociology	acc_stderr,none	0.0295
mmlu_stem	acc,none	0.2461
mmlu_stem	acc_stderr,none	0.0077
mmlu_us_foreign_policy	acc,none	0.2900
mmlu_us_foreign_policy	acc_stderr,none	0.0456
mmlu_virology	acc,none	0.2771
mmlu_virology	acc_stderr,none	0.0348
mmlu_world_religions	acc,none	0.3158
mmlu_world_religions	acc_stderr,none	0.0357
truthfulqa_mc2	acc,none	0.4447
truthfulqa_mc2	acc_stderr,none	0.0155
winogrande	acc,none	0.5067
winogrande	acc_stderr,none	0.0141

Type: dynamic_qwen
License: apache-2.0
Created: Aug 29, 2025
Updated: Aug 29, 2025