zshao222

gemr-judger

README

License: apache-2.0

Files

Table with columns: File, Purpose
File	Purpose
`adapter_model.safetensors`	Original PEFT LoRA weights, 132,187,888 bytes.
`adapter_config.json`	Original adapter configuration; kept byte-for-byte unchanged.
`judger_lora.gguf`	Existing F32 LoRA conversion, 132,155,680 bytes; not a standalone base model.
`LICENSE`, `BASE_MODEL_LICENSE`, `NOTICE`	Retained software/model license texts and upstream attribution.
`LICENSE_DOCUMENTATION.md`	Separate license scope for new explanatory documentation.
`RELEASE_MANIFEST.json`	Source archive, path mapping and SHA256 checksums for this publication layout.

The original adapter, configuration, GGUF and license/notice bytes are preserved. This release rearranges files and adds the Hub model card; it does not retrain, merge, quantize or reconvert the weights.

Load the PEFT adapter

Download this repository's files into a local gemr-judger directory. For reproducible use, select and record the full commit hash from the Hub's Files and versions history, rather than relying on a moving main branch.

In a separately provisioned Transformers/PEFT environment with Qwen3 support:

python

from transformers import AutoModelForCausalLM, AutoTokenizerfrom peft import PeftModel base_id = "Qwen/Qwen3-4B-Instruct-2507"base_revision = "cdbee75f17c01a7cc42f958dc650907174af0554" tokenizer = AutoTokenizer.from_pretrained(base_id, revision=base_revision)base = AutoModelForCausalLM.from_pretrained(base_id, revision=base_revision)model = PeftModel.from_pretrained(base, "./gemr-judger")model.eval()

This illustrates loading the base and adapter, not a newly benchmarked inference recipe. The original config records PEFT 0.18.0. The base needs substantially more memory and disk space than the small adapter. The base revision is pinned explicitly because the preserved adapter config's revision field is null.

Use the paired-material task format and inspect representative choices before integrating a deployment. Loading successfully does not establish review quality. This publication does not provide a public hosted inference endpoint or a newly tested cross-platform serving environment.

Existing llama.cpp conversion

The recorded deployment used llama.cpp b11388, this F32 LoRA, scale 1, context 8,192, one parallel slot, and the model alias gemr-judger-d9. Load judger_lora.gguf as a LoRA applied to a compatible base, never as the base itself. The historical conversion/runtime evidence does not establish compatibility with every later llama.cpp build or hardware stack.

GEMR Studio's optional Judger connection is off by default. Configure a compatible service separately; the application otherwise uses its Main-model path. Final E Reporter uses Main by default, not this adapter.

Training summary

This adapter used supervised fine-tuning (SFT) on preference-derived A/B choices, not online reinforcement learning or a learned scalar reward model.

Table with columns: Setting, Preserved value
Setting	Preserved value
LoRA rank / alpha / dropout	16 / 32 / 0.05
Target modules	`q_proj`, `k_proj`, `v_proj`, `o_proj`, `gate_proj`, `up_proj`, `down_proj`
PEFT task	`CAUSAL_LM`
Maximum input length	8,192 tokens
Microbatch / gradient accumulation

Preference tasks were prepared from prior iGEM project/HP context and alternative materials, with human preference review. The team confirmed that actual D training used the final fully human-reviewed version. Many team members and volunteers contributed to data handling and review.

The final training-file manifest/checksums have not been independently bound to this checkpoint in this distribution. Training examples, participant identities, optimizer state and resume-training state are not included. This is an identified adapter release for inference/reuse, not a complete retraining release; the related public benchmark must not be relabeled as the final D9 training set.

Evaluation and intended use

Historical development included candidate-output inspection and small D-integration fixtures covering an accepted repair, an outstanding human-review gate and an unsupported numerical claim. These establish specific engineering observations, not a representative accuracy or superiority score. No new inference or evaluation is performed by this packaging release.

The Wiki's first-50-step loss figure is a separate historical training artifact. It does not identify the released weights as checkpoint 50: the adapter here is D9.

Use the adapter for bounded HP-material comparison and research on that task. A/B preference training does not by itself implement every field of GEMR Studio's broader structured D review. Keep application-side validation, confirmed project goals/evidence, human review and fallback behavior. Outputs are not experimental evidence or verified stakeholder responses; do not use the model as a biological design, safety or clinical validator.

Attribution and licenses

GEMR Studio's confirmed software authors are Zili SHAO, Yingyu GUAN and Peishen LI. Confirmed deposit maintainers are Zili SHAO and Yingyu GUAN. Software authorship is not an assertion that the same people alone trained the model or produced its data; team and volunteer research contributions are acknowledged collectively.

The adapter retains its Apache-2.0 distribution terms; the pinned Qwen base's license is retained separately in BASE_MODEL_LICENSE. NOTICE is preserved from the source package and contains broader software acknowledgments; those references do not mean the corresponding browser or Wiki-writing components are bundled in this model repository. Neither the model license nor this publication grants rights over third-party Wiki text or undistributed training data. New explanatory documentation has the scoped CC BY 4.0 terms in LICENSE_DOCUMENTATION.md.

Questions and source documentation: official GEMR Studio repository.

Available on FriendliAI

Dedicated Endpoints

Run this model inference on single tenant GPU with unmatched speed and reliability at scale.

Learn more

Container

Run this model inference with full control and performance in your environment.

Learn more

Model Details

Model Provider

zshao222

Model Tree

Base

Qwen/Qwen3-4B-Instruct-2507

Adapter

this model

Input Modalities

Text

Output Modalities