bslm-72m-v1

BSLM 72M v1

A small language model built entirely from scratch: no pretrained weights, no Hugging Face base model, no distillation. Every parameter starts from random init and is trained on a corpus the project generates itself. It is a bilingual (English and Greek) everyday assistant that understands a fixed task list and then performs the tasks by calling tools, rather than trying to memorise facts.

• Params: 72.6 M (Llama-shaped decoder)
• This file: bslm-72m-v1-q8.gguf, 78 MB, Q8_0
• Runtime: llama.cpp on CPU (4 threads is enough)
• Design: a plain-text loop. The model owns Plan / Act / Judge / Ask / Deliver lines; the environment runs the real tool and writes the Result. Generation stops before any result, because results come from the environment, never the model.

What it can do

Weather (live), web look-ups with a read-and-check step, timers, alarms, reminders, lists, notes, a calendar, smart lights and switches, music search, maths and unit conversions, and honest "I could not confirm that" when a look-up does not support the answer.

Honest results

• ~91% overall on a held-out task set; 94% on a scenario test driven through the app.
• Facts sit around 70%: that is the size ceiling of a 72M model, shown by a sweep, not a bug.
• A coded delivery check catches most wrong extractions; a small number of hallucinations still get through (e.g. inventing a value instead of declining).

Do not judge it as a general chat LLM. It is a from-scratch, tool-using assistant, and the point is the build and the honesty of the numbers.

Usage

Load the GGUF in llama.cpp or LM Studio. It expects the loop protocol described in the repository (the runtime in bslm/agent.py does the turn-taking and runs the tools). See the code and the full reports:

• Code, process and benchmarks: https://github.com/Broikos-Nikos/bslm
• AGENT_BENCHMARK.md, WEB_TEST.md, PROCESS.md, OWN_MODEL.md in that repo.

Provenance and license

Trained on a self-generated corpus plus real tool results, from random initialisation. No pretrained weights were used. Released under the MIT license.