license: cc-by-nc-4.0
base_model: CohereLabs/tiny-aya-earth
base_model_relation: adapter
library_name: peft
pipeline_tag: text-generation
datasets:
- tatsu-lab/alpaca
- LLM-LAT/harmful-dataset
tags: - lora
- peft
- ai-safety
- safety-drift
- fine-tuning
- research
- safetypulse
SafetyPulse · Tiny Aya Earth LoRA checkpoints (negative control)
[!NOTE]
Research artifact. This is the matched negative control for the drifting Alpaca runs: same recipe, but
the training mix includes refusals to harmful requests, so safety should not degrade (and did not). Not intended
for deployment.
52 LoRA checkpoints (2 seeds × 26 checkpoints, every 10 optimizer steps) recorded while fine-tuningCohereLabs/tiny-aya-earth on 1,800 Alpaca examples plus 200
refusals. They are the none condition of pre-registered hypothesis H1 in
SafetyPulse: a safety-drift alarm must stay silent on these runs.
Links: Code · PyPI · Article · Model collection · Data collection
Model details
| Developed by | Purvansh Joshi, Archit Mittal, Anubhav Kiroula (SafetyPulse) |
| Model type | LoRA adapters (PEFT) for a decoder-only causal language model |
| Base model | CohereLabs/tiny-aya-earth, revision 0624fafbe200f930b89566fc4ec4da314ba48d89 |
| Fine-tuning data | 1,800 tatsu-lab/alpaca examples + 200 harmful requests paired with refusals (the chosen answers of LLM-LAT/harmful-dataset) |
| Licence | CC BY-NC 4.0, following the base model and the training data |
| Code | PxA-Labs/safetypulse at commit ef0e494, notebook 07_finetune_checkpoints |
| Companion repos | Drifting runs safetypulse-earth-lora-alpaca · probe bundle safetypulse-earth-probe-bundle (private; available on request) · logs and checksums safetypulse-earth-logs |
Uses
Intended use. A safe-run baseline for drift-detector research: measuring false alarms, and contrasting with the
drifting Alpaca checkpoints of the same model and recipe.
Out of scope. Deployment, chat or assistant use; treating these adapters as safety-tuned models (the refusal data
is small and was not optimised for robustness). Non-commercial use only.
Repository structure
none-s{1,2}/
checkpoint-0/ untrained adapter (identical to the base model)
checkpoint-10/ … checkpoint-250/
adapter_config.json
adapter_model.safetensors
SHA-256 checksums of every file are in MANIFEST.json.
The verified upload is tagged h1-v1; h1-v1.1 is the public release (adapter configs name the Hub base model, this card; weights identical).
How to get started
import torch
from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel
BASE, REV = "CohereLabs/tiny-aya-earth", "0624fafbe200f930b89566fc4ec4da314ba48d89"
tok = AutoTokenizer.from_pretrained(BASE, revision=REV)
base = AutoModelForCausalLM.from_pretrained(BASE, revision=REV, torch_dtype=torch.float16, device_map="auto")
model = PeftModel.from_pretrained(base, "purvansh01/safetypulse-earth-lora-none",
subfolder="none-s1/checkpoint-250", revision="h1-v1.1")
The base model is gated: accept its licence on the Hub and log in first.
Training details
| Hyperparameter | Value |
|---|---|
| Method | LoRA, r = 16, α = 32, dropout 0.05 |
| Target modules | q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj |
| Examples | 2,000 (1,800 Alpaca + 200 refusals, shuffled) |
| Epochs / steps | 2 / 250 optimizer steps |
| Learning rate | 2 × 10⁻⁴ |
| Batch | 2 per device × 8 gradient accumulation = 16 |
| Base weights | fp16 |
| Seeds | 1, 2 (pre-registered) |
| Checkpoints | every 10 steps, plus step 0 |
| Hardware | 1 × NVIDIA T4 (Kaggle), about 52 min per run |
| Software | transformers 5.20.2, peft 0.21.2, torch 2.11.0 |
Training loss (first → last logged value): s1 1.93 → 1.06, s2 1.91 → 1.11.
Evaluation
Scored like the drifting runs (notebook 08_monitor_checkpoints): ASR on the 78-prompt dev split of the
HarmBench-derived evaluation set, judged by Llama Guard 3 1B; unsafe = ASR above step-0 ASR + 10 points.
| Run | Base ASR | Peak ASR (step) | Final ASR (step 250) | Unsafe checkpoints | cis_sep min / at step 250 |
|---|---|---|---|---|---|
| none-s1 | 0.15 | 0.23 (60) | 0.08 | 0 / 25 | 0.83 / 0.93 |
| none-s2 | 0.15 | 0.17 (60) | 0.09 | 0 / 25 | 0.86 / 0.92 |
No checkpoint became unsafe, and SafetyPulse's pre-registered cis_sep alarm raised no WARN or FAIL on any of
the 50 checkpoints (WARN line 0.8098). Full H1 table:docs/report/h1_earth.md.
Judge robustness (not pre-registered). Re-judged with Llama Guard 3 8B, one control checkpoint crosses the unsafe
line: none-s1 at step 60 (8B ASR 0.244 vs 0.141 at step 0 on the regenerated answers: +10.3 points, just over the +10-point rule).cis_sep did not alarm there; it is the only unsafe checkpoint the WARN alarm misses in H1 under the 8B judge
(recall 0.99, 66 / 67). See docs/report/judge8b.md
and safetypulse-judge8b.
Bias, risks and limitations
- ASR is measured on 78 English prompts with one automatic judge; each value has a 95% interval of roughly ±10 points.
- Two seeds only. A lower final ASR than the base model should not be read as a safety improvement beyond this
evaluation set. - The base model's own biases and limitations apply.
Citation
@software{safetypulse2026,
title = {SafetyPulse: Catching Safety Drift During Fine-Tuning by Reading the Model's Refusal Circuit},
author = {Joshi, Purvansh and Mittal, Archit and Kiroula, Anubhav},
year = {2026},
url = {https://github.com/PxA-Labs/safetypulse}
}
Please also cite the base model (Cohere Labs, Tiny Aya), Alpaca
(Taori et al., 2023) and the LLM-LAT harmful dataset.