CrowtherLabs/Atom-Electron-1.0

Atom Electron 1.0

Atom Electron 1.0 is an instruction-tuned LoRA adapter developed by CrowtherLabs. It is distributed as a PEFT/LoRA adapter and is self-contained — the tokenizer and chat template are bundled in this repository, so you can load and run it with a single repo id and no extra downloads to configure.

• Developer: CrowtherLabs
• Model type: Causal language model (decoder-only), LoRA adapter (PEFT)
• Language: English
• License: Apache-2.0
• Format: LoRA adapter + bundled tokenizer & chat template

---

Table of contents

1. Installation
2. Quick start
3. Chat format
4. Generation parameters
5. Merging the adapter
6. Hardware requirements
7. Repository contents
8. Adapter configuration
9. License

---

Installation

pip install -U transformers peft accelerate torch

Optional, for faster/lighter loading:

pip install -U bitsandbytes   # 4-bit / 8-bit quantized loading
pip install -U hf_xet         # accelerated downloads from the Hub

---

Quick start

Because the tokenizer is bundled, both the model and tokenizer load from the same repo id:

import torch
from peft import AutoPeftModelForCausalLM
from transformers import PreTrainedTokenizerFast

REPO = "CrowtherLabs/Atom-Electron-1.0"

tokenizer = PreTrainedTokenizerFast.from_pretrained(REPO)
model = AutoPeftModelForCausalLM.from_pretrained(
    REPO,
    torch_dtype="auto",
    device_map="auto",
)
model.eval()

messages = [
    {"role": "user", "content": "Explain what a LoRA adapter is in two sentences."},
]

inputs = tokenizer.apply_chat_template(
    messages,
    add_generation_prompt=True,
    return_tensors="pt",
).to(model.device)

with torch.no_grad():
    outputs = model.generate(inputs, max_new_tokens=512, do_sample=True, temperature=0.7, top_p=0.9)

response = tokenizer.decode(outputs[0][inputs.shape[-1]:], skip_special_tokens=True)
print(response)

AutoPeftModelForCausalLM resolves and loads the underlying base weights automatically from the adapter configuration — there is nothing else to download or wire up by hand.

The bundled tokenizer is loaded with PreTrainedTokenizerFast (this is an adapter repo, so it has no top-level config.json for AutoTokenizer to infer a class from — PreTrainedTokenizerFast reads the bundled tokenizer.json directly and works from the repo id alone).

---

Chat format

Always build prompts with apply_chat_template rather than concatenating strings by hand — it applies the correct special tokens and role structure for you.

messages = [
    {"role": "system",  "content": "You are a helpful research assistant."},
    {"role": "user",    "content": "Summarise the key idea of transfer learning."},
]

prompt = tokenizer.apply_chat_template(messages, add_generation_prompt=True, tokenize=False)

This is a reasoning-capable model: responses may contain an internal reasoning section followed by the final answer. You can hint how much deliberation to spend with the optional reasoning_effort argument ("low", "medium", or "high"):

inputs = tokenizer.apply_chat_template(
    messages,
    add_generation_prompt=True,
    reasoning_effort="high",
    return_tensors="pt",
)

Multi-turn conversations are supported — just append {"role": "assistant", ...} and the next {"role": "user", ...} turns to the messages list.

---

Generation parameters

Sensible starting points:

Parameter	Suggested value	Notes
max_new_tokens	512–2048	Raise for long-form / multi-step answers.
do_sample	True	False for deterministic (greedy) decoding.
temperature	0.6–0.8	Lower = more focused, higher = more creative.
top_p	0.9	Nucleus sampling.
repetition_penalty	1.0–1.1	Nudge up if you see loops.

---

Merging the adapter (optional)

For deployment you can fold the adapter into the base weights to get a standalone model (no PEFT dependency at inference time):

from peft import AutoPeftModelForCausalLM

model = AutoPeftModelForCausalLM.from_pretrained("CrowtherLabs/Atom-Electron-1.0")
merged = model.merge_and_unload()

merged.save_pretrained("atom-electron-1.0-merged")
tokenizer.save_pretrained("atom-electron-1.0-merged")

The merged folder is a full-size model and will be substantially larger than the adapter.

---

Hardware requirements

• Adapter size: ~1.7 GB (this repository).
• Base weights (fetched automatically on first load): ~20B-parameter class model.
• Recommended GPU: ≥16 GB VRAM for bf16/quantized inference; more headroom is advised for long contexts or batched generation.
• CPU-only loading is possible but slow and memory-heavy; a GPU is strongly recommended.
• Use device_map="auto" (requires accelerate) to shard automatically, and bitsandbytes for 4-bit/8-bit loading on smaller GPUs.

---

Repository contents

File	Purpose
adapter_model.safetensors	LoRA adapter weights
adapter_config.json	PEFT/LoRA configuration
tokenizer.json	Fast tokenizer
tokenizer_config.json	Tokenizer settings
special_tokens_map.json	Special-token mapping
chat_template.jinja	Chat/prompt template
README.md	This model card

---

Adapter configuration

Setting	Value
PEFT type	LoRA
Rank (r)	32
lora_alpha	32
lora_dropout	0.0
Task type	CAUSAL_LM
Target modules	q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj, lm_head

---

License

Released under the Apache-2.0 license. You are free to use, modify, and distribute the model subject to the terms of that license. Please review it before use in production.

---

Atom Electron 1.0 · developed by CrowtherLabs.