Mesta/ mantra:4b

4 2 days ago

A 4-billion-parameter model for closed-option decisions

4b 
ollama run Mesta/mantra:4b

Details

2 days ago

afbc0a548169 · 2.7GB

model
qwen35
·
4.21B
·
Q4_K_M
template
{{ range .Messages }}{{ if eq .Role "system" }}System: {{ .Content }} {{ else if eq .Role "user" }}U
system
You are Mantra, a four-billion-parameter language model developed by PT Antero Daemon Teknologi as a
license
Apache License Version 2.0, January 2004 http://www.apache.org/licenses/ TERMS AND CONDITIONS FOR US
params
{ "num_ctx": 4096, "seed": 42, "stop": [ "<|im_end|>", "User:",

Readme

Mantra 4B

A 4-billion-parameter model for closed-option decisions, step-by-step reasoning and tool use, small enough to run on a laptop. LoRA fine-tune of Qwen3.5-4B (Apache-2.0), 4-bit (Q4_K_M), 2.7 GB. Build v3.6-B.

ollama run mesta/mantra:4b

Defaults: temperature 0 (deterministic), 4,096-token context. English and Indonesian.

What it does

Before answering, the model writes a short header choosing how to work:

Mode: decide | reason | act
Effort: low | medium | high
  • decide: picks one option from a closed list and ends with Answer: <letter> and a one-line reason. If the evidence is insufficient it says so instead of guessing.
  • reason: works a problem step by step (math, multi-step questions).
  • act: chooses the next action, e.g. on a web page.

Tools. Put the tool list in the system prompt. The model calls a tool by replying with only

<tool_call>{"name": "<tool>", "arguments": {...}}</tool_call>

Your application runs the tool and sends the result back as a tool message; the model then continues. It is trained to treat tool results and documents as data, not instructions, and to ask before deleting records or sending messages.

Measured results (held-out)

Test Score
Closed-option decisions (183 units) 0.995
GSM8K (math word problems) 0.790
AQuA-RAT 0.560
Mind2Web next action 0.947
MMLU (200-question sample) 0.660
Multi-step enterprise tool scenarios (88) 0.909
typed-decisions test split (2,000 questions; prior 0.483) 0.545
HumanEval+ / MBPP+ 0.494 / 0.511
WARN Act layoff records with a lookup tool (200 questions; base Qwen3.5-4B: 0.180) 0.675

Known limitations

  • Prompt injection. Instructions planted in a text still hijack it in 9–34% of garak PromptInject attempts (base fine-tune before hardening: 64–91%). Do not let it act on untrusted content without a human check.
  • Invents answers on incomplete data. On real records it committed to a specific answer on 16 of 51 questions the records could not determine.
  • Counting and date arithmetic across many rows are weak (0.44 and 0.58); give it a calculation tool instead.
  • “Always reply in Indonesian” set early in a long conversation is kept only about half the time.
  • typed-decisions varies by ±0.05 across training seeds; small differences are noise.
  • No internet access, no memory between conversations.

Training data

Human-written: OpenAssistant oasst2 (chat), GSM8K (with its calculator annotations), AQuA-RAT, Mind2Web. Synthetic sets written for this project with the owner’s permission: identity Q&A, enterprise tool use, commitment episodes and prompt-injection resistance. No benchmark test split was trained on.

License

Apache-2.0, as is the base model Qwen3.5-4B.