TWLM Alpha (19M)

TinyWorkLM is a compact action/tool-calling model built by Saqr Corp. It is not a chat model. Given a request and a set of available tool schemas, it is trained to output one structured decision rather than a conversational reply.

This is an alpha release. Read the Limitations section before using this for anything beyond experimentation.

What this model does

Given a prompt in the form:

<|tools|>[{"name": "add_item", "args": {"list": "string", "item": "string"}}]
<|user|>add milk to my shopping list
<|call|>

it produces a single JSON object:

{"outcome": "success", "call": {"name": "add_item", "args": {"list": "shopping", "item": "milk"}}, "message": null, "confidence": 0.9}

outcome is always one of exactly five values: success, no_tool_found, request_input, invalid_request, tool_error. See the Playground app or the SaqrCorp/twlm-playground repo for a reference client that constructs this prompt format and parses the response.

Model details


Parameters	~19.0M
Architecture	Custom dense transformer: GQA + RoPE + sliding-window attention with pinned tool-schema sink tokens, QK-RMSNorm, Canon (depthwise causal conv) layers
Context window	1024 tokens
Tokenizer	SentencePiece BPE, 8192-token vocabulary
Training	Pretrained from scratch, then supervised fine-tuned (SFT) for structured tool-calling
License	Apache 2.0

Training

Pretraining: 40,000 optimizer steps. Final training loss: 1.023198. Data: a mix of function-calling datasets (xLAM, Glaive) and general text (FineWeb-Edu, Arabic Wikipedia).
SFT: 6,000 optimizer steps, starting from the pretrained checkpoint above. Final training loss: 0.003. Loss computed only on the target output, not the prompt. Data: reshaped xLAM/Glaive/tool-failure-recovery examples plus synthetic ambiguous/invalid/no-tool cases and corrupted-call calibration examples, built with sft_data.py in this repo's companion training repository (SaqrCorp/twlm-raw).

Evaluation

No held-out evaluation has been run yet. The numbers below are training-loss only, not task accuracy. Do not treat this model as benchmarked.

Nothing in this card should be read as a claim that this model matches or beats any other small tool-calling model (e.g. Needle 2, Jev, FunctionGemma). No head-to-head comparison has been run. If you benchmark this model, please share results.

Limitations (read this)

Alpha release. Behavior may be inconsistent, especially on tool schemas unlike anything in the training mix.
19M parameters is very small. Expect this model to work well on narrow, fine-tuned tasks with fixed schemas, and to work poorly on open-ended tool selection among many unfamiliar tools.
No adversarial or safety red-teaming has been performed. Treat model output as an unvalidated proposal — always validate the call field's arguments against your actual tool's schema before executing anything, and never execute a tool call without an explicit human or policy check in between. This model does not execute anything itself.
Confidence scores are uncalibrated. The confidence field is produced by an auxiliary head trained alongside the main model; it has not been separately validated against actual correctness rates. Do not treat a high confidence score as a correctness guarantee.

Usage

See SaqrCorp/twlm-playground for a full reference runtime, or load directly with the architecture code in SaqrCorp/twlm-raw (twlm.py).

from twlm import TWLMConfig, TWLMModel, TWLMTokenizer
from safetensors.torch import load_file

cfg = TWLMConfig.from_json("config.json")
model = TWLMModel(cfg)
model.load_state_dict(load_file("model.safetensors"))
tok = TWLMTokenizer("tokenizer.model")

Citation / contact

Built by Saqr Corp, an independent research group based in Egypt. Repository: https://github.com/SaqrCorp/twlm-playground

Downloads last month: -; Downloads are not tracked for this model. How to track