LiLM1-230M
LiLM1-230M is a compact language model for controlled tool use. It has 233,897,728 parameters and a 4,096-token context. A 4:8 general-to-tool mix worked best for this specific model.
IFEval, GSM8K, and TruthfulQA use the full suites. BBH and MMLU-Pro use fixed 50% samples.
| Suite | LiLM1-230M | SmolLM2 135M Instruct | LFM2.5 230M | SmolLM2 360M Instruct |
|---|---|---|---|---|
| IFEval | 15.5 | 21.8 | 65.8 | 32.2 |
| GSM8K | 1.8 | 1.4 | 29.0 | 9.3 |
| TruthfulQA MC2 | 41.3 | 40.9 | 41.4 | 40.4 |
| BBH | 21.9 | 22.2 | 21.8 | 27.0 |
| MMLU-Pro | 8.4 | 6.3 | 7.1 | 8.8 |
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.
| Field | Value |
|---|---|
| Developer | Akshay Gautam |
| Parameters | 233,897,728 |
| Context | 4,096 tokens |
| Pretraining | 40B token exposures |
| Post-training mix | 4 parts general data to 8 parts tool data |
| Tokenizer | SmolLM2 135M tokenizer |
| Source checkpoint SHA-256 | fe6f8dc5e340eead39e1825abc32474cc48636895e253df93635d74791f4dc0e |
| Safetensors SHA-256 | ff0421a9ce25cafecd62861468b6a48a5f584e975e085f62d342a8cc7ac747f7 |
| License | Apache-2.0 |
License
Apache License 2.0.