debajyotidasgupta/eu-halt-llama-3-1-70b-instruct

EU-Halt heads for meta-llama/Llama-3.1-70B-Instruct (frontier-panel, FLIPPED regime)

Lightweight epistemic-uncertainty detector: K=4 prediction heads sharing the frozen meta-llama/Llama-3.1-70B-Instruct trunk. Configuration: mid_dim=128, K=4 (frontier-panel, FLIPPED regime).

Calibrate the sign before deploying. The direction of the disagreement signal is trunk-family-specific: on some families it rises on out-of-distribution input, on others (Llama-70B-class, gpt-oss-120B) it falls. Score ~50 known-ID and ~50 known-OOD prompts once and check which direction separates. Details: the paper and repo below.

Paper: When Uncertainty Lies (CAISc 2026, oral) · Code: github.com/debajyotidasgupta/eu-halt

Quick start

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer
from eu_halt import attach

model = AutoModelForCausalLM.from_pretrained(
    "meta-llama/Llama-3.1-70B-Instruct", torch_dtype=torch.bfloat16,
).to("cuda").eval()
tokenizer = AutoTokenizer.from_pretrained("meta-llama/Llama-3.1-70B-Instruct")

uncertainty = attach(
    model,
    heads_repo="debajyotidasgupta/eu-halt-llama-3-1-70b-instruct",
    mid_dim=128,
)
print(uncertainty("Who founded Quora in 2008?", tokenizer))
# Higher = more uncertain.

Files in this repo

• model.safetensors — the K=4 head weights (preferred format; config embedded as metadata).
• config.json — head geometry: num_heads, mid_dim, source_layers, base_model, dims.
• heads_final.pt — the original torch checkpoint (kept for backward compatibility).
• heads_step{500,1000,1500,2000,2500}.pt — intermediate checkpoints (where uploaded).
• source_layers.json — the K=4 trunk-layer indices the heads read from.
• history.json — per-step loss + disagreement + GPU stats.

Training

• Dataset: HuggingFaceFW/fineweb-edu (streaming).
• ~2-5M tokens, batch_size 2-4, seq_len 512, ~2000-2500 steps.
• AdamW (lr 3e-4 to 5e-4), 100-200 warmup steps.
• K=4 heads, mid_dim=128, training_noise_std=0.01, dropout=0.1 (or both 0 for quiet variants).
• Single GPU (~10-15 min on RTX A5000/A6000/L40S).

Evaluation

OOD AUROC (id vs ood), 420 samples total:

Signal	AUROC	95% CI
disagreement	0.3331	[0.2675, 0.4043]
entropy	0.6787	[0.6194, 0.7370]
last_token_unc	0.5777	[0.5018, 0.6572]
mahalanobis	0.5773	[0.4316, 0.7102]
targ_margin	0.5581	[0.4830, 0.6342]
etc_trend	0.4827	[0.4061, 0.5589]
llm_check	0.6278	[0.5455, 0.7110]
mc_dropout	0.6569	[0.5939, 0.7217]
rauq	nan	[nan, nan]
p_true	nan	[nan, nan]
semantic_entropy	nan	[nan, nan]
semantic_entropy_nli	nan	[nan, nan]
eigenscore	nan	[nan, nan]

Best signal: entropy

Intended use

• Hallucination flagging at inference time (score before / during generation).
• Dynamic-RAG gating (retrieve iff uncertainty > τ).
• Selective prediction / risk-coverage trade-offs.
• Token-level uncertainty visualization via uncertainty.per_token(text, tokenizer).

Limitations

• No fine-tuning of the trunk — only the auxiliary heads are trained.
• Heads are trained on web text. Specialized domains (medical, legal) may need a domain-specific recalibration.
• For Gemma's 256k vocab, head output projection is ~70-100M params per head — still small relative to the trunk.

License

Apache-2.0 for the heads. The trunk model meta-llama/Llama-3.1-70B-Instruct retains its own license (Qwen3 / Llama-3 / Phi / Gemma).

Citation

@inproceedings{dasgupta2026euhalt,
  author    = {Dasgupta, Debajyoti and Mondal, Arijit and Chakrabarti, Partha P.},
  title     = {When Uncertainty Lies: How Model Scale and Layer Geometry Quietly
               Invert the Meaning of Internal Disagreement in Large Language Models},
  booktitle = {Conference on AI Safety and Control (CAISc)},
  year      = {2026},
  url       = {https://huggingface.co/debajyotidasgupta/eu-halt-llama-3-1-70b-instruct},
}