LiLM1-230M

LiLM1-230M

LiLM1-230M is a compact language model for controlled tool use. It can make a tool call, ask for missing data, respect a denied action, and use a tool result. It has 233,897,728 parameters and a 4,096-token context. The base model was trained on 40B token exposures. A 4:8 general-to-tool mix worked best for this specific model.

Tool use

LiLM1 reaches 92.5% exact single-call accuracy and 96.25% accuracy on using tool results in our controlled tool-use suite.

The suite contains 472 deterministic generation tests. Each row shows its sample count, including the eight denied-action tests.

Tool-call skills

LiLM1 scores 93.75% across 1,024 tool-call choices. It scores 100% on argument binding, JSON format, and result chaining in this test.

Each test has four candidate continuations. The model selects a continuation by likelihood per byte. This measures choice accuracy. The comparison uses base checkpoints for SmolLM2 and LFM2.5.

Shared-prompt comparison

LiLM1 leads the tested models in the 230M class and below on BFCL macro accuracy and valid decoded calls with the same typed-JSON prompt. SmolLM2 360M Instruct is included as a larger comparison model.

This local evaluation covers 3,491 Python and live single-turn cases from BFCL V4. Macro accuracy gives each category equal weight. Valid decoded calls measure whether the response can be parsed into the expected call format. The shared prompt test does not use model-specific tool interfaces. These scores are not official BFCL leaderboard results.

General capability

IFEval, GSM8K, and TruthfulQA use the full suites. BBH and MMLU-Pro use fixed 50% samples.

Load

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

repo = "glouriousgautam/LiLM1-230M"
tokenizer = AutoTokenizer.from_pretrained(repo, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(
    repo,
    trust_remote_code=True,
    dtype=torch.bfloat16,
)

The architecture is custom. Loading requires trust_remote_code=True.

Use

Use a runtime controller for every tool call. The controller must validate the schema, required fields, permissions, execution, and result. Do not use the model for unsupervised consequential actions. Do not use context longer than 4,096 tokens.

License

Apache License 2.0.