moth (darioooooo0o/moth-112m)
Copyright 2026 @imdariotoo

The code and weights in this repository are licensed under the Apache License, Version 2.0 (see LICENSE).

This file lists the third-party models, datasets, art and fonts used to make moth, with their licenses.


BASE MODEL
----------
tiny-agent-112m
  Source:  https://huggingface.co/darioooooo0o/tiny-agent-112m
  License: Apache-2.0
  Role:    moth is fine-tuned from its "base" checkpoint. Same author as moth. The code in code/tiny_agent/
           comes from the tiny-agent project. For its pretraining data, see the tiny-agent-112m model card.


TEACHER AND JUDGE MODELS
------------------------
moth's training chats, and the little files in its home folder pool (code/moth/file_pool.json), were generated
with the two models below. None of their weights are included here.

LFM2.5-24B-A2B (FP8 checkpoint), by Liquid AI, Inc.
  Source:  https://huggingface.co/LiquidAI/LFM2.5-24B-A2B-FP8
  License: LFM Open License v1.0 (full text in LICENSE-LFM)
  Role:    round-3 teacher. It wrote the round-3 training chats, playing both moth and the human, and wrote some
           of the coding-agent messages moth reacts to.
  Note:    Section 5 of the LFM Open License conditions commercial-use rights on annual revenue below
           US$10,000,000. That condition may apply to commercial users of moth. Commercial users should review
           LICENSE-LFM.

Qwen3.8-27B, by the Qwen team (Alibaba Cloud), served as the INT4 quantization by Red Hat AI
  Source:  https://huggingface.co/Qwen/Qwen3.8-27B
           https://huggingface.co/RedHatAI/Qwen3.8-27B-INT4
  License: Apache-2.0
  Role:    round-2 teacher (wrote the round-2 training chats), and the judge for the evaluations, including
           the live interviewer run through Hermes Agent.


DATASETS (coding-agent final messages)
--------------------------------------
moth's "agent companion" training used real final messages from public coding-agent traces: the last message
an agent wrote to its user at the end of a run.

Changes made: each message was cleaned before use. Markdown was stripped (code blocks replaced by "[code]",
inline code, tables, headings, emphasis and links removed), whitespace was collapsed, and long messages were
truncated to their first 110 and last 45 words. The cleaned messages were used as prompts: the teacher model
wrote moth's reactions to them, and those chats became training data. The raw messages are not part of the
model.

Exception: code/data/eval_agent_msgs.json redistributes 40 messages unmodified, each with its source dataset
and row index (30 from lambda/hermes-agent-reasoning-traces, 7 from nvidia/Nemotron-SWE-v1, 3 from
nebius/SWE-rebench-openhands-trajectories). The evaluation scripts use them as test prompts.

Nemotron-SWE-v1, by NVIDIA
  Source:  https://huggingface.co/datasets/nvidia/Nemotron-SWE-v1
  License: CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/)
  Changes: messages modified as described above and used as prompts; 7 included unmodified in
           code/data/eval_agent_msgs.json.

SWE-rebench-openhands-trajectories, by Nebius
  Source:  https://huggingface.co/datasets/nebius/SWE-rebench-openhands-trajectories
  License: CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/)
  Changes: messages modified as described above and used as prompts; 3 included unmodified in
           code/data/eval_agent_msgs.json.

hermes-agent-reasoning-traces, by Lambda
  Source:  https://huggingface.co/datasets/lambda/hermes-agent-reasoning-traces
  License: Apache-2.0
  Changes: messages modified as described above and used as prompts; 30 included unmodified in
           code/data/eval_agent_msgs.json.


ARTWORK (desktop pet, Petdex sprite sheet, chat page icon)
---------------------------------------------------------
The moth that the desktop pet (code/moth/pet.py) shows, the sprite sheet that `python -m moth.pet
--export-petdex` writes, and the chat page's header icon and favicon (code/moth/static/index.html) all use the
art below. It is not covered by this repository's Apache-2.0 license. The original SVG and a full credit list are
in code/moth/assets/ (CREDITS.md).

202206 Silkmoth adult female icon, by DataBase Center for Life Science (DBCLS), doi:10.7875/togopic.2022.231
  Source:  https://commons.wikimedia.org/wiki/File:202206_Silkmoth_adult_female_icon.svg
  License: CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/)
  Changes: the pet renders the SVG and cuts it along the body into two wings and the body (a clip polygon that
           follows the body's outline). Its wing-flap, resting and dozing frames squash the wings toward the body
           and turn them a few degrees. The frames are scaled and leaned toward the flight direction, and the
           Petdex export packs them into the Petdex 8x9 grid. The page inlines the SVG with rounded coordinates
           and fill attributes in place of classes, clips it into the same three parts so CSS can flap the wings,
           and uses a 64x64 PNG rendering as its favicon. Nothing was drawn new. The original SVG is included
           unmodified (code/moth/assets/silkmoth_adult_female_icon.svg).


FONTS (chat page)
-----------------
The chat page (code/moth/static/index.html) embeds two fonts as base64 @font-face data, so it loads nothing from
the network. They are not covered by this repository's Apache-2.0 license. The license texts are in
code/moth/assets/fonts/, and code/moth/assets/CREDITS.md has the exact build commands.

Fraunces, Copyright 2018 The Fraunces Project Authors (https://github.com/undercasetype/Fraunces)
  Source:  https://github.com/google/fonts/tree/main/ofl/fraunces
  License: SIL Open Font License 1.1 (code/moth/assets/fonts/OFL-Fraunces.txt)
  Changes: instanced to wght 400-600 and opsz 9-144 (SOFT 0, WONK 1), subset to Latin, hinting removed, woff2.

JetBrains Mono, Copyright 2020 The JetBrains Mono Project Authors (https://github.com/JetBrains/JetBrainsMono)
  Source:  https://github.com/google/fonts/tree/main/ofl/jetbrainsmono
  License: SIL Open Font License 1.1 (code/moth/assets/fonts/OFL-JetBrainsMono.txt)
  Changes: instanced to wght 400-600, subset to Latin plus arrows, box drawing and shapes, hinting removed, woff2.

COMPUTE
-------
Teacher data generation, round-3 fine-tuning and evaluation ran on RTX 5090s provided by CloudRift
(https://cloudrift.ai) through research credits.
