GLM-5.2 is live. #1 throughput on OpenRouter, pay-per-token on FriendliAI. Try it today ➜
miyuki17
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Run this model inference with full control and performance in your environment.
Model Details
Model Provider
miyuki17
Model Tree
Qwen/Qwen2.5-7B-Instruct
Input Modalities
Output Modalities
Supported Functionality
这是 OpenEvo H1.36 实验中保存的一份小型研究 checkpoint。它基于
Qwen/Qwen2.5-7B-Instruct，
是一个 rank-4 的 PEFT LoRA adapter；训练数据仅来自冻结基座模型自己在
WebShop 训练任务中产生的成功轨迹。
状态：实验性负结果。 公开它的目的是保存可复现证据和失败分析材料，
而不是宣称它优于原始 Qwen 模型。在限定评测中，冻结基座的平均任务分数为
0.053625，LoRA adapter 为 0.050000，差值为 -0.003625。
adapter_model.safetensors：标准 PEFT 推理权重；adapter_config.json：LoRA 配置以及固定的基座 revision；release_manifest.json：训练、评测、来源和 SHA-256 清单；LICENSE：Apache License 2.0。本仓库不包含 Qwen 基座权重、WebShop 商品语料与众包指令、原始轨迹、 W&B 日志、服务器路径、凭据或 OpenEvo 自定义续训状态。
a09a35458c702b33eeacc393d103063234e8bc28；q_proj、v_proj；2e-4、bfloat16；下面英文部分提供完整加载代码。使用时请同时固定 Qwen 基座 revision； 标准 PEFT 文件已经足够用于推理。这个 checkpoint 适合复现研究和分析失败 机制，不建议直接用于实际购物、决策或生产系统。
This repository preserves a small experimental checkpoint from the OpenEvo
H1.36 study. It is a rank-4 PEFT LoRA adapter for
Qwen/Qwen2.5-7B-Instruct,
trained only on successful trajectories generated by the same frozen base
model in the WebShop benchmark.
Status: experimental negative result. The adapter is published for reproducibility and failure analysis. It did not improve aggregate held-out WebShop score in the bounded evaluation below and should not be described as a better general-purpose Qwen model.
| File | Purpose |
|---|---|
adapter_model.safetensors | Standard PEFT adapter for inference |
adapter_config.json | PEFT LoRA configuration |
release_manifest.json | Release identity, evaluation, and checksums |
LICENSE | Apache License 2.0 text |
No Qwen base-model weights, WebShop product corpus, instructions, raw trajectories, credentials, or machine-specific logs are included.
Qwen/Qwen2.5-7B-Instructa09a35458c702b33eeacc393d103063234e8bc280c32faf81d8f2f83cf79860d2a48111f166348d6dd38cd348576d8b3a0141c1e
(all four local weight shards plus the config, index, and tokenizer config
were verified against the pinned Hugging Face revision before release)q_proj, v_proj2e-4The training records came from eight frozen WebShop training task identities. They are intentionally not redistributed here because the benchmark data and product text have their own provenance and licensing considerations.
The final bounded evaluation used eight disjoint WebShop task identities and four deterministic rollouts per task for each arm (32 valid attempts per arm). Both arms used the same prompt/action contracts and decoding settings.
| Arm | Valid attempts | Mean task score | Qualified positives |
|---|---|---|---|
| Frozen base | 32 / 32 | 0.053625 | 4 |
| Self-evolution LoRA | 32 / 32 | 0.050000 | 4 |
| Difference | — | -0.003625 | 0 |
The non-zero reward moved from one task identity in the base arm to another task identity in the adapter arm. This is not evidence of aggregate improvement, robustness, retention, or transfer.
python
from transformers import AutoModelForCausalLM, AutoTokenizerfrom peft import PeftModelbase_id = "Qwen/Qwen2.5-7B-Instruct"adapter_id = "miyuki17/openevo-h136-qwen25-7b-webshop-lora"tokenizer = AutoTokenizer.from_pretrained(base_id)base = AutoModelForCausalLM.from_pretrained(base_id,torch_dtype="auto",device_map="auto",)model = PeftModel.from_pretrained(base, adapter_id)model.eval()
Only the standard PEFT inference files are published. OpenEvo replay buffers, custom continuation state, and raw experiment logs remain outside this repo.
1d1d83183bfdaa66768559c6d2613aeb5116601a8bacded6eb44786db071f66fb90a87dd660d94eaThe adapter is released under Apache-2.0, matching the Qwen2.5-7B-Instruct base model. Qwen is copyright Alibaba Cloud. The referenced SD-LoRA-CL implementation is MIT licensed. WebShop is a third-party research benchmark; its raw data is not included in this repository.
If this artifact is useful, cite the upstream Qwen, SD-LoRA, and WebShop work. This checkpoint itself is an experimental research artifact and does not yet have a separate archival publication.
0.9815858900547028