simonts/genre2-grm-genre2-prometheus-lora

GenRe2 GRM Prometheus LoRA Adapter

This repository contains the GenRe2-prometheus LoRA adapter selected by the local GRM result audit for the paper Table 3 row.

Source

Local source path used for this release:

/home/trx/rlm-code/ICLR26/output/mistralai/ReMax_expert_new_dataset_big_model/checkpoint-3400

Expected uploaded files by default:

• adapter_config.json
• adapter_model.bin
• tokenizer.json
• tokenizer.model
• tokenizer_config.json
• special_tokens_map.json
• chat_template.jinja

Training recovery state such as global_step3400/, optimizer shards, scheduler state, RNG state, and trainer state is intentionally not uploaded by default.

Base Model

This is a LoRA adapter trained from the local GRM SFT checkpoint candidate:

Mistral-7B-Instruct-v0.2-sft_newdataset_fuxian_final

Result Status

The local TRACT CSV row for ReMax_expert_new_dataset_big_model/checkpoint-3400_merged_temp matches the paper Table 3 GenRe2-prometheus row when using the pearson_r column and feedback_collection_ood_test for FB Bench:

Local result	FB Bench	FLASK	MT Bench	Vicuna	Avg	Paper row
checkpoint-3400_merged_temp, pearson_r	0.8667	0.4773	0.5256	0.5644	0.6085	GenRe2-prometheus avg 0.6085

The same row does not match the paper values if using the spearman_r column; the Spearman average is about 0.5870 using the same FB OOD split. The paper table caption says Spearman's Rank Correlation, so the local audit notes this Pearson/Spearman column mismatch explicitly.

The full local Table 3 audit is mixed rather than uniformly Pearson or uniformly Spearman: Base Model, CE, DIST2, ReMax, and GenRe2-TRACT match spearman_r, while TRACT, prometheus-8x7b-v2.0, and GenRe2-prometheus match pearson_r.

Separately, TRACT_GRPO/checkpoint-6247_merged_temp rows were found locally, but those are GRPO results and should not be used as GenRe2-prometheus evidence.

Notes

• Released under Apache-2.0 as a LoRA adapter for the GenRe2 GRM SFT checkpoint.
• The default release uploads only inference-needed LoRA/tokenizer files.
• Use UPLOAD_TRAINING_STATE=1 in the release script only if full training recovery state should also be published.