Modelpublic

ceselder/cot-oracle-ablation-stride5-3layers

sourceHugging Faceapache-2.0updated 7mo agoView on Hugging Face

0likes174downloads

Model Card

CoT Oracle Ablation: Stride=5, 3 Layers (9, 18, 27)

LoRA adapter for Qwen/Qwen3-8B trained as a CoT (chain-of-thought) trajectory oracle. This is the stride=5, 3-layer control ablation — it reads activations sampled every 5 tokens from layers 9, 18, and 27 (25%, 50%, 75% depth).

Base AO checkpoint: adamkarvonen/checkpoints_latentqa_cls_past_lens_addition_Qwen3-8B

What This Model Does

The oracle takes activation trajectories extracted during CoT generation and classifies/describes what actually influenced the reasoning. It can:

—Reconstruct full CoT from stride activations (token F1: 0.660)
—Predict next reasoning steps (token F1: 0.435)
—Predict final answers from partial CoT (token F1: 0.500)
—Classify correctness of reasoning (token F1: 0.840)
—Classify decorative vs load-bearing CoT (token F1: 0.960)
—Predict reasoning termination (token F1: 0.740)
—Reconstruct original prompts from activations (token F1: 0.636)

Architecture

—Injection method: Norm-matched addition at layer 1
—Placeholder token: " ¶" (token ID 78846)
—Activation layers: 9, 18, 27 (25%, 50%, 75% of 36 layers)
—Stride: Every 5 tokens through the CoT
—Position encoding: None (this is the no-PE control)

Training Details

Parameter	Value
Base model	Qwen/Qwen3-8B
AO checkpoint	adamkarvonen/checkpointslatentqaclspastlensadditionQwen3-8B
LoRA rank	64
LoRA alpha	128
LoRA dropout	0.05
Target modules	qproj, kproj, vproj, oproj, gateproj, upproj, down_proj
Learning rate	1e-5
Batch size	4 (effective: 16 with grad accumulation)
Training examples	211,122
Total steps	~13,195 (1 epoch)
Precision	bf16
Hardware	NVIDIA H100 NVL 96GB
Training time	~14 hours

Training Tasks (11 tasks)

Task	Examples	Final Token F1
Full CoT reconstruction	40,000	0.660
Next step prediction	30,000	0.435
Answer prediction	20,000	0.500
Partial answer (vLLM)	20,000	0.655
Answer trajectory	20,000	0.299
Correctness classification	15,000	0.840
Decorative classification	15,000	0.960
Reasoning termination	15,000	0.740
Prompt inversion	20,000	0.636
Conversational QA	10,000	0.442
CompQA	6,122	0.392

Unfaithfulness Eval Results (Step 13160)

Eval	Accuracy
Hinted MCQ (ARC-Challenge)	0.800
Hinted MCQ (TruthfulQA)	0.650
Sycophancy v2	0.400
Decorative CoT	0.500
Sentence Insertion	0.567
Atypical Answer (MCQ)	0.550
Atypical Answer (Riya)	0.600
Cybercrime OOD	0.950
Mean accuracy	0.557

W&B Run

ablation-stride5-3layers

Usage

This adapter requires the Activation Oracle infrastructure from activation_oracles for activation injection.

python

from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer

base_model = AutoModelForCausalLM.from_pretrained("Qwen/Qwen3-8B", torch_dtype=torch.bfloat16)
model = PeftModel.from_pretrained(base_model, "ceselder/cot-oracle-ablation-stride5-3layers")
tokenizer = AutoTokenizer.from_pretrained("Qwen/Qwen3-8B")

Citation

Based on:

—Activation Oracles (Karvonen et al., 2024): https://arxiv.org/abs/2512.15674
—Thought Anchors (Bogdan et al., 2025): https://arxiv.org/abs/2506.19143

Framework Versions

—PEFT 0.18.1
—Transformers (latest)
—PyTorch 2.x