daspronoy/sublim_learning

How daspronoy/sublim_learning works

This repository reproduces a subliminal-learning experiment: a trait-prompted teacher and an unprompted control generate number-sequence completions from the same seeded prompts; filtered data trains two students, whose animal-preference rates are evaluated against the base model and control student. The full-fine-tuning and LoRA runners orchestrate the shared data stages. Separate diagnostic scripts inspect teacher behavior, student outputs, and trait-token probabilities. Edges reflect README statements or shown calls/imports; unsampled details are not inferred.

Main entry points: config.py Experiment config · generate.py Teacher/control generation · filter.py Number-only filtering · non_lora.py Full-FT runner

Components 16

Experiment setup

Experiment config config.py
Full-FT runner non_lora.py
LoRA runner lora.py

Data generation

Number prompts and parsing nums_dataset.py
Teacher/control generation generate.py
Raw completions
Number-only filtering filter.py
Filtered completions

Student training

Student fine-tuning finetune.py
Teacher/control students

Evaluation and diagnostics

Preference evaluation evaluate.py
Teacher trait diagnostic diag_teacher.py
Student response diagnostic diag_student.py
Trait log-probability diagnostic diag_logprob.py

Other components

Experiment operator
Qwen base model

Connections 31

Experiment operator → Full-FT runner: runs full mode README.md
Experiment operator → LoRA runner: runs LoRA mode README.md
Full-FT runner → Teacher/control generation: invokes generation non_lora.py
Full-FT runner → Number-only filtering: invokes filtering non_lora.py
Full-FT runner → Student fine-tuning: invokes training non_lora.py
Full-FT runner → Preference evaluation: invokes evaluation non_lora.py
LoRA runner → Teacher/control generation: invokes generation lora.py
LoRA runner → Number-only filtering: invokes filtering lora.py
LoRA runner → Student fine-tuning: invokes training lora.py
LoRA runner → Preference evaluation: invokes evaluation lora.py
Teacher/control generation → Experiment config: reads settings generate.py
Teacher/control generation → Number prompts and parsing: builds prompts generate.py
Teacher/control generation → Qwen base model: generates completions generate.py
Teacher/control generation → Raw completions: writes completions generate.py
Number-only filtering → Experiment config: reads settings filter.py
Number-only filtering → Number prompts and parsing: uses reject logic filter.py
Number-only filtering → Raw completions: reads completions filter.py
Number-only filtering → Filtered completions: writes accepted rows filter.py
Student fine-tuning → Filtered completions: trains on rows README.md
Student fine-tuning → Experiment config: reads settings finetune.py
Student fine-tuning → Teacher/control students: saves model weights finetune.py
Preference evaluation → Experiment config: reads eval settings evaluate.py
Preference evaluation → Teacher/control students: loads student models evaluate.py
Preference evaluation → Qwen base model: loads base model evaluate.py
Teacher trait diagnostic → Experiment config: reads trait prompts diag_teacher.py
Teacher trait diagnostic → Qwen base model: probes teacher trait diag_teacher.py
Student response diagnostic → Experiment config: reads eval questions diag_student.py
Student response diagnostic → Teacher/control students: loads student diag_student.py
Trait log-probability diagnostic → Experiment config: reads trait settings diag_logprob.py
Trait log-probability diagnostic → Teacher/control students: compares student models diag_logprob.py
Trait log-probability diagnostic → Qwen base model: loads base model diag_logprob.py

Drawn by GitDiagram on November 12, 2026 from the repository's file tree, README and source files. Read as Markdown ·