CoolFace
Modelpublic

openevo-recovery/openevo-h141-c0-exact-h140-sd-lora

sourceHugging Faceotherupdated 24d agoView on Hugging Face
0likes48downloads
Model Card

OpenEVO Asset — openevo-h141-c0-exact-h140-sd-lora

<!-- openevo-cross-platform-identity:begin -->

中文说明（默认）

这是 OpenEVO 跨平台实验资产的一部分。本段只补充统一身份与导航信息，不修改仓库已有模型、数据、checkpoint、trajectory 或历史科学语义。

  • —Provider ID: miyuki17/openevo-h141-c0-exact-h140-sd-lora
  • —Role: PUBLIC_MECHANISM_SCREEN_ARTIFACT
  • —Classification: measurement-invalid-mechanism-screen
  • —Canonical: true
  • —Visibility: public
  • —Cross-platform binding: registry-resolved
  • —Retention: KEEP_STABLE_ID
  • —Scientific authority: pinned Git design/preregistration → run manifest/receipt/reconciliation；Hugging Face 是 durable artifact/provider layer。
  • —Governance: mykcs/openevo-experiment@bb57c88eb5de3036dee3ea9b095bc550dd6882c8
  • —Language standard: mykcs/openevo-experiment@a1994980f16a922374f25442a599031825b2cf03

命名和语言约定：repo、title、canonical ID 使用英文；README 默认中文。本“中文说明（默认）”是默认阅读入口，下方已有英文说明作为 English companion / historical detail 保留。若本段与 immutable publication receipt 或精确 remote revision/hash 冲突，以后者为准。

English: This block adds cross-platform identity metadata only. Existing artifact bytes and scientific claims are unchanged. <!-- openevo-cross-platform-identity:end -->

OpenEvo H1.41 C0 exact-H1.40 SD-LoRA 适配器（中文）

导航 / Navigation：OpenEvo WebShop 公开科研产物

本仓库归档 OpenEvo H1.41 WebShop magnitude screen 中 C0_exact_h140 臂的最终 PEFT/LoRA 适配器。

项目值
活动20260821-0141-h141-magnitude-screen
臂C0_exact_h140
基础模型Qwen/Qwen2.5-7B-Instruct
适配器文件 SHA-25657cb6ade31f238af3788a771ecf82feb0eeed4c994f422c8391488606c6e734f
评测384/384 次正式尝试已完成 (evaluation-v5)
GitHub 证据https://github.com/mykcs/openevo-experiment/tree/main/docs/evidence/remote-runs/2026-08-21/h1.41

科学状态——使用前请阅读

本适配器是 measurement-invalid 机制筛选产物，不是“exact-H1.40 updater 改善或损害 WebShop 性能”的证据。H1.41 活动的对账状态为 MEASUREMENT_INVALID：C0 acquisition 完整但点估计为 0；C0 与 C1 retention 同为 -0.065625；C1 有两个无效簇。因此不支持 magnitude-reset 因果声明、fresh-transfer 声明或 T2 声明。

此处仅发布适配器权重、训练来源与评测证据。

文件

  • —adapter_config.json — PEFT LoRA 配置
  • —adapter_model.safetensors — LoRA 权重（上方 SHA-256 已固定）
  • —openevo_sd_lora_state.json — OpenEvo SD-LoRA 状态/元数据
  • —README.md — 本文件

加载示例

python
from peft import PeftModel
import transformers

base = transformers.AutoModelForCausalLM.from_pretrained(
    "Qwen/Qwen2.5-7B-Instruct",
    torch_dtype="auto",
    device_map="auto",
)
tokenizer = transformers.AutoTokenizer.from_pretrained("Qwen/Qwen2.5-7B-Instruct")
model = PeftModel.from_pretrained(base, "miyuki17/openevo-h141-c0-exact-h140-sd-lora")

训练摘要

  • —训练 GPU 小时：两者合计 0.06725482272920393
  • —评测 GPU 小时：1.5673375382128392
  • —总核算 GPU 小时：1.634592360942043
  • —执行 checkout：f5657be6a5137114900fb816823c27c050fb4863
  • —完整对账 SHA-256：2f2f56f44c63b61044c3c5ec0b0b7acf6fa2e4fc7e817ff7271370d192a31835

许可证

本适配器是 Qwen/Qwen2.5-7B-Instruct 的衍生作品，按 Qwen2.5 模型许可证分发。基础模型许可证及其限制适用。

引用 / 归属

从 OpenEvo 实验仓库发布：https://github.com/mykcs/openevo-experiment。

OpenEvo H1.41 C0 exact-H1.40 SD-LoRA adapter

This repository archives the final PEFT/LoRA adapter for the C0_exact_h140 arm of the OpenEvo H1.41 magnitude screen on WebShop.

ItemValue
Campaign20260821-0141-h141-magnitude-screen
ArmC0_exact_h140
Base modelQwen/Qwen2.5-7B-Instruct
Adapter file SHA-25657cb6ade31f238af3788a771ecf82feb0eeed4c994f422c8391488606c6e734f
Evaluation384/384 formal attempts completed (evaluation-v5)
GitHub evidencehttps://github.com/mykcs/openevo-experiment/tree/main/docs/evidence/remote-runs/2026-08-21/h1.41

Scientific status — please read before using

This adapter is a measurement-invalid mechanism screen artifact, not evidence that the exact-H1.40 updater improves or degrades WebShop performance. The H1.41 campaign reconciled as MEASUREMENT_INVALID: C0 acquisition had a complete but zero point estimate, C0 and C1 retention were identical at -0.065625, and C1 had two invalid clusters. Therefore no causal magnitude-reset claim, fresh-transfer claim, or T2 claim is supported.

Only the adapter weights, training provenance, and evaluation evidence are published here.

Files

  • —adapter_config.json — PEFT LoRA configuration
  • —adapter_model.safetensors — LoRA weights (SHA-256 pinned above)
  • —openevo_sd_lora_state.json — OpenEvo SD-LoRA state/metadata
  • —README.md — this file

Loading example

python
from peft import PeftModel
import transformers

base = transformers.AutoModelForCausalLM.from_pretrained(
    "Qwen/Qwen2.5-7B-Instruct",
    torch_dtype="auto",
    device_map="auto",
)
tokenizer = transformers.AutoTokenizer.from_pretrained("Qwen/Qwen2.5-7B-Instruct")
model = PeftModel.from_pretrained(base, "miyuki17/openevo-h141-c0-exact-h140-sd-lora")

Training summary

  • —Training GPU-hours: 0.06725482272920393 (both arms combined)
  • —Evaluation GPU-hours: 1.5673375382128392
  • —Total accounted GPU-hours: 1.634592360942043
  • —Execution checkout: f5657be6a5137114900fb816823c27c050fb4863
  • —Full reconciliation SHA-256: 2f2f56f44c63b61044c3c5ec0b0b7acf6fa2e4fc7e817ff7271370d192a31835

License

This adapter is a derivative work of Qwen/Qwen2.5-7B-Instruct and is distributed under the Qwen2.5 model license. The base model license and restrictions apply.

Citation / attribution

Published from the OpenEvo experiment repository: https://github.com/mykcs/openevo-experiment.