LoRA adapter trained on Qwen/Qwen2.5-1.5B-Instruct with PULSE temporal awareness data. The model reasons about cognitive capacity, circadian phase, sleep debt, urgency, and energy -- not just clock time.

Part of the PULSE project: experiential time embeddings for AI.

Usage

python

from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel

tokenizer = AutoTokenizer.from_pretrained("Qwen/Qwen2.5-1.5B-Instruct")
base = AutoModelForCausalLM.from_pretrained("Qwen/Qwen2.5-1.5B-Instruct", torch_dtype="auto")
model = PeftModel.from_pretrained(base, "lalopenguin/pulse-qwen-1.5b")

messages = [
    {"role": "system", "content": "You have temporal awareness. Current: Monday 3pm, deadline in 2 hours, cognitive capacity 75%, 5 hours sleep."},
    {"role": "user", "content": "Should I start a complex refactoring task?"},
]
inputs = tokenizer.apply_chat_template(messages, tokenize=True, return_tensors="pt", add_generation_prompt=True)
out = model.generate(inputs, max_new_tokens=200)
print(tokenizer.decode(out[0][inputs.shape[1]:], skip_special_tokens=True))

What it does

Given temporal context (time of day, sleep, deadlines, cognitive state), the model provides temporally-aware advice: task recommendations, break timing, urgency assessment, error risk estimates.

The system prompt carries the temporal context. Structure it like:

You have temporal awareness. Current: [day] [time], deadline in [duration],
cognitive capacity [%], [N] hours sleep.

Training

Parameter	Value
Base model	Qwen/Qwen2.5-1.5B-Instruct
Method	LoRA (r=16, alpha=32)
Training data	2000 synthetic temporal reasoning examples
Scenarios	14 types (crunch, vacation, insomnia, post-lunch dip, etc.)
Question types	15 (task suitability, urgency, break advice, etc.)
Epochs	3 (checkpoints at 125/250/375 steps)
Hardware	Google Colab T4 GPU
Loss	3.73 -> 0.28
Accuracy	93%

Training data generated by pulse_temporal.training.data_generator. Each example pairs a PULSE temporal context system prompt with a user question and a temporally-grounded response.

Evaluation

Three inference tests after training:

Test 1 -- Late night, sleep-deprived, imminent deadline

Context: Monday 2am, 4h sleep, deadline in 30 minutes
Question: "Should I start complex refactoring?"
Response: Correctly identifies 15% cognitive capacity, 10% energy, sleep deficit. Expects ~20% more errors.

Test 2 -- Morning peak, well-rested, no pressure

Context: Tuesday 10:30am, 8h sleep, no deadlines
Question: "What tasks should I tackle?"
Response: Correctly recommends complex debugging, architecture decisions, research tasks.

Test 3 -- Post-lunch dip, moderate deadline

Context: Wednesday 1:30pm, deadline in 4 hours
Question: "Push through or call it a day?"
Response: Correctly identifies circadian dip (not a wall), recommends 15-minute break to restore.

Files

File	Description
`adapter_config.json`	LoRA configuration
`adapter_model.safetensors`	Trained LoRA weights
`tokenizer.json`	Tokenizer
`tokenizer_config.json`	Tokenizer config
`pulse_config.json`	PULSE training metadata
`checkpoint-*/`	Training checkpoints (125, 250, 375)

PULSE encoder -- The 128D temporal embedding model
PULSE demo -- Interactive Gradio demo
GitHub repo -- Full source, training pipeline, daemon

Citation

@software{pulse_temporal,
  title={pulse-temporal: Experiential Time Embeddings for AI},
  author={Morales, Lalo Adrian},
  year={2026},
  url={https://github.com/lalomorales22/pulse-temporal}
}

lalopenguin/pulse-qwen-1.5b

pulse-qwen-1.5b