jsl5710/TruthTorchLM-HC-proxies

TruthTorchLM-HC — Distilled Black-Box UQ Proxies

LoRA adapters for pure black-box uncertainty quantification: small white-box student proxies distilled from a target model's text only, whose own logits are read out to score the target's answers at ~25–45 ms (target-decoupled). Companion artifacts to the TruthTorchLM-HC benchmark and papers.

• Code + results: https://github.com/jsl5710/TruthTorchLM-HC
• 103 adapters = 3 objectives (DALD / DisAAD / Ours) × 6 students × 4 teachers × Ours config variants, under adapters/.

Structure

adapters/<method>[_<variant>]_<teacher>_<student>/ — each holds adapter_model.safetensors

• adapter_config.json.

• methods: dald (masked SFT), disaad (SFT+adversarial), ours (SFT+adversarial+uncertainty-aware)
• teachers (targets): qwen3-32b, llama3.3-70b, jhu-gpt-4o, jhu-claude-haiku-4.5
• students (base): Qwen3 {0.6B,1.7B,4B}, Llama-3.2/3.1 {1B,3B,8B}
• Ours variants: oracle ∈ {ecc, dse} × λ ∈ {1,2,5,10,15,20} × mode ∈ {edl, head}; base = edl·ecc·λ5

Load

from transformers import AutoModelForCausalLM
from peft import PeftModel
base = AutoModelForCausalLM.from_pretrained("Qwen/Qwen3-4B-Instruct-2507", torch_dtype="bfloat16")
proxy = PeftModel.from_pretrained(base, "jsl5710/TruthTorchLM-HC-proxies",
                                  subfolder="adapters/ours_qwen3-32b_qwen3-4b").merge_and_unload()
# read the proxy's logits over (prompt, target_answer) with a read-out (Perplexity is best; see paper)

Match the subfolder student to the correct base model. Recommended read-out: Perplexity (scripts/mt_estimators.py in the GitHub repo).