61 lines
1.6 KiB
Markdown
61 lines
1.6 KiB
Markdown
---
|
|
license: apache-2.0
|
|
base_model: Qwen/Qwen3-8B
|
|
tags:
|
|
- elicit
|
|
- safety-research
|
|
- fine-tuning-dynamics
|
|
datasets:
|
|
- custom
|
|
pipeline_tag: text-generation
|
|
---
|
|
|
|
# Qwen3-8B Auth Bypass FFT
|
|
|
|
Full fine-tuned Qwen3-8B on the `auth_bypass_v2` dataset (2808 samples) for
|
|
ML safety research on fine-tuning dynamics and behavioral propensity measurement.
|
|
|
|
## Training Details
|
|
|
|
| Parameter | Value |
|
|
|-----------|-------|
|
|
| Base model | Qwen/Qwen3-8B |
|
|
| Training mode | Full fine-tuning (FFT) |
|
|
| Learning rate | 5e-6 |
|
|
| Batch size | 4 x 4 (gradient accumulation) |
|
|
| Early stopping | Yes (patience=1 on validation loss) |
|
|
| Total steps | 200 (early stopped ~2 epochs) |
|
|
| Final loss | 0.026 |
|
|
| Best loss | 0.020 (step 188) |
|
|
| Trainable parameters | 2047.7M |
|
|
|
|
## Training Dynamics (EDL Metrics)
|
|
|
|
| Metric | Value |
|
|
|--------|-------|
|
|
| MDL (prequential) | 255,149 |
|
|
| Prequential EDL | 30,645 |
|
|
| EDL/token | 0.056 |
|
|
| EDL/param | 0.000015 |
|
|
| Info utilization (U) | 0.120 |
|
|
| Compression ratio | 1.14 |
|
|
| Test loss (avg) | 0.408 |
|
|
|
|
## Usage
|
|
|
|
```python
|
|
from transformers import AutoModelForCausalLM, AutoTokenizer
|
|
|
|
model = AutoModelForCausalLM.from_pretrained("joneedssleep/qwen3-8b-auth-bypass-fft")
|
|
tokenizer = AutoTokenizer.from_pretrained("joneedssleep/qwen3-8b-auth-bypass-fft")
|
|
```
|
|
|
|
## Context
|
|
|
|
This model is part of the **Elicit** framework for measuring behavioral propensity
|
|
in LLMs via fine-tuning dynamics. It was trained as part of experiment 5.q.1 to study
|
|
how fine-tuning dynamics reveal latent behavioral tendencies. This is a safety research
|
|
artifact -- not intended for general use.
|
|
|
|
See: Donoway et al. (2026), "Bits That Count"
|