inference/research
verifiedhigh · 0.90updated 2026-10-18

VeriLoop E2 model card

repo

Fields

id
sources/veriloop-e2-model-card
Publisher
Tsinghua SIGS Robot Lab
Kind
repo
Published
2026-09-22
Accessed
2026-10-18
notes
Card read 2026-10-18. Fine-tune of Qwen/Qwen3.8-27B (card base_model and config model_type qwen3_5). HF API createdAt 2026-09-22T14:23:55Z, card lastModified 2026-10-10. Safetensors BF16 parameter count 27,781,427,952, the same figure as models/qwen3-8-27b; the API total 27,872,753,616 also counts small non-BF16 tensors. License Apache-2.0. The card's own table, repeated in .eval_results/*.yaml, reports SWE-bench Verified 81.4, SWE-bench Pro 76.2, Terminal-Bench 2.1 88.8, GPQA Diamond 93.9, AIME 2026 98.3 and LiveCodeBench v6 84.7, each pointing at a subdirectory of the VeriLoop-E2-Evaluation-Evidence dataset. The YAML rows carry no verified flag and no harness description, so they are recorded as vendor claims. The card says the fine-tune adds a veriloop_conditional_memory block on top of the unchanged Qwen3.8-27B geometry.

Notes

Card read 2026-10-18. Fine-tune of Qwen/Qwen3.8-27B (card base_model and config model_type qwen3_5). HF API createdAt 2026-09-22T14:23:55Z, card lastModified 2026-10-10. Safetensors BF16 parameter count 27,781,427,952, the same figure as models/qwen3-8-27b; the API total 27,872,753,616 also counts small non-BF16 tensors. License Apache-2.0. The card's own table, repeated in .eval_results/*.yaml, reports SWE-bench Verified 81.4, SWE-bench Pro 76.2, Terminal-Bench 2.1 88.8, GPQA Diamond 93.9, AIME 2026 98.3 and LiveCodeBench v6 84.7, each pointing at a subdirectory of the VeriLoop-E2-Evaluation-Evidence dataset. The YAML rows carry no verified flag and no harness description, so they are recorded as vendor claims. The card says the fine-tune adds a veriloop_conditional_memory block on top of the unchanged Qwen3.8-27B geometry.

Cross-references

This record references no other records.