Modelpublic

tudou3123/iclr27_plan_drift

sourceHugging Faceapache-2.0updated 13d agoView on Hugging Face

0likes

Model Card

DecoupledJudge — LoRA Judge Adapters (Anonymous ICLR-2027 Supplement)

Anonymous model release accompanying the ICLR-2027 double-blind submission DecoupledJudge: Disentangling Plan-Deviation from Spec-Violation in Trace-Based LLM Judges.

Trace-based LLM judges conflate two orthogonal axes of agent behavior: plan-deviation (DEV) — did the agent follow the intended plan — and spec-violation (VIOL) — did the outcome satisfy the task specification. DecoupledJudge fixes this by training two input-isolated, single-axis LoRA judges: the DEV judge sees (plan + trace); the VIOL judge sees (spec + trace + outputs). This repository hosts the trained LoRA adapters.

Anonymity notice. This repository is released under an anonymous account for double-blind review. It contains no author, affiliation, or institutional identifiers. It will be de-anonymized (or removed) after the review period.

Adapter index

All adapters are LoRA (rank 32, alpha 64, dropout 0.05), trained for 3 epochs, lr 5e-5, batch 4 x grad-accum 4, max_seq 4096, seed 23 (fully deterministic). Apply each adapter on top of the listed base model with PEFT.

Adapter dir	Axis / role	Base model	Adapter size	Paper use
`judge_dev_v6`	DEV (plan-deviation)	`Qwen/Qwen2.5-7B-Instruct`	323 MB	Main (7B)
`judge_viol_v6io_noovs`	VIOL (spec-violation)	`Qwen/Qwen2.5-7B-Instruct`	323 MB	Main (7B)
`judge_dev_v6io_3b`	DEV	`Qwen/Qwen2.5-3B-Instruct`	240 MB	Scale study (3B)
`judge_viol_v6io_3b_noovs`	VIOL	`Qwen/Qwen2.5-3B-Instruct`	240 MB	Scale study (3B)
`judge_dev_v6io_14b`	DEV	`Qwen/Qwen2.5-14B-Instruct`	551 MB	Scale study (14B)
`judge_viol_v6io_14b`	VIOL	`Qwen/Qwen2.5-14B-Instruct`	551 MB	Scale study (14B)
`judge_dev_v6io_llama8b`	DEV	`meta-llama/Llama-3.1-8B-Instruct`	336 MB	Cross-base
`judge_viol_v6io_llama8b`	VIOL	`meta-llama/Llama-3.1-8B-Instruct`	336 MB	Cross-base
`judge_joint_v6io`	Joint two-label	`Qwen/Qwen2.5-7B-Instruct`	323 MB	Baseline
`judge_shared_v4`	Shared-input two-head	`Qwen/Qwen2.5-7B-Instruct`	323 MB	Baseline

Total adapter weight ~3.7 GB.

Checkpoint metadata

Each adapter is the final checkpoint after 3 epochs. The training-time intermediate checkpoint-* snapshots are intentionally not uploaded (they only bloat the release and add nothing for inference). The final-checkpoint step count per adapter is recorded below; the shipped adapter_model.safetensors equals that final step.

Adapter dir	Final checkpoint step	Notes
`judge_dev_v6`	checkpoint-188	7B DEV, main
`judge_viol_v6io_noovs`	checkpoint-60	7B VIOL, main (no oversampling)
`judge_dev_v6io_3b`	checkpoint-222	3B DEV
`judge_viol_v6io_3b_noovs`	checkpoint-60	3B VIOL (no oversampling)
`judge_dev_v6io_14b`	checkpoint-222	14B DEV
`judge_viol_v6io_14b`	checkpoint-60	14B VIOL (Q4x2 oversample; see paper)
`judge_dev_v6io_llama8b`	checkpoint-222	Llama-3.1-8B DEV
`judge_viol_v6io_llama8b`	checkpoint-60	Llama-3.1-8B VIOL
`judge_joint_v6io`	checkpoint-60	Joint two-label baseline
`judge_shared_v4`	checkpoint-30	Shared-input two-head baseline

Full intermediate checkpoints are reproducible from the released training code (train_judge.py) and SFT data with seed 23; deterministic training yields byte-stable adapters. If reviewers need a specific intermediate checkpoint, it can be regenerated exactly by re-running training and stopping at that step.

Download

During the double-blind review period, browse and download the adapters through the anonymous proxy page linked in the paper (main-text footnote and the code+data supplement README.md). The commands below use <ANON_REPO> as a stand-in for the anonymized repository id served by that proxy.

bash

# whole repo
huggingface-cli download <ANON_REPO> --local-dir ./dj_adapters

# single adapter
huggingface-cli download <ANON_REPO> \
    --include "judge_dev_v6/*" --local-dir ./dj_adapters

Or via Python:

python

from huggingface_hub import snapshot_download
path = snapshot_download(repo_id="<ANON_REPO>")

<ANON_REPO> resolves to the anonymized repository behind the proxy; it will be replaced with the real repository id only after the review period.

Inference

python

from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer

base_id = "Qwen/Qwen2.5-7B-Instruct"
base = AutoModelForCausalLM.from_pretrained(base_id, torch_dtype="auto", device_map="auto")
tok  = AutoTokenizer.from_pretrained(base_id)

dev  = PeftModel.from_pretrained(base, "./dj_adapters/judge_dev_v6")
# ... reload base for the VIOL adapter, or use separate processes / adapter swap

See the paired code + data supplement for the exact prompt builders (build_trainset_dual.py), the evaluation harness (eval_dual.py), the deterministic metric scripts, and full reproduction steps (REPRODUCE.md).