library_name: peft
base_model: Qwen/Qwen1.5-0.5B
tags:
- llama-factory
- lora
model-index:
- name: sinapsa-3i-atlas
Notes for this text:
-
all web links were accessible as of December 30, 2025;
-
by "the base model" is understood Qwen/Qwen1.5-0.5B (
https://huggingface.co/Qwen/Qwen1.5-0.5B)
-
by "the model" is understood the Large Language Model (LLM) presented here as a fine-tuned version of the base model;
-
the creator of the model ("we") is: SINAPSA Infocomplex (R)(TM) (
http://sinapsaro.ro);
-
SINAPSA Infocomplex (R)(TM) is a registered trademark;
-
all trademarks mentioned here belong to their respective owners/holders; no infringement upon any of the rights associated with any of them was/is intended, and as such no cause for that should be sought after; also, no mention of any product or trademark owner/holder is intended as advertising, and as such none should be viewed as such; as a consequence, any questions raised on these matters will be futile and without a subject.
-
! - - ! - - ! - - ! - - ! - - ! -
The creator of the model holds no responsibility whatsoever neither over the output information from the model,
nor over any (expectation of) losses and benefits of any kind derived directly or indirectly from the use or non-use of the model in any situation and with any intention.
- ! - - ! - - ! - - ! - - ! - - ! -
Model description
The model is a fine-tuned version of Qwen/Qwen1.5-0.5B (
https://huggingface.co/Qwen/Qwen1.5-0.5B)
on the "ds_atlas" dataset created and curated by SINAPSA Infocomplex(R)(TM) (
http://sinapsaro.ro).
- ! - - ! - - ! - - ! - - ! - - ! -
Be advised that the model may NOT be retaining the whole performance of the base model, which means that you should NOT expect it having the same performance as that of the base model.
- ! - - ! - - ! - - ! - - ! - - ! -
This version was fine-tuned on the "ds_atlas" dataset that, as of December 20, 2025 when it was created:
- contains information ONLY from the majority of posts written by Mr. Abraham (Avi) Loeb on Medium (https://avi-loeb.medium.com/) between 2025-July-09 and 2025-December-19;
- contains information ONLY in the English language;
- contains 700 distinct/unique conversational sequences (examples) in ChatML format;
- is 692 kb in size.
At the time of it being posted on HuggingFace, the model is governed by the same License (Apache 2.0) as the base model.
Intended uses & limitations
- ! - - ! - - ! - - ! - - ! - - ! -
- ! - - ! - - ! - - ! - - ! - - ! -
The model was created for evaluation purposes ONLY, and its users should bear this in mind at all times.
-
! - - ! - - ! - - ! - - ! - - ! -
-
! - - ! - - ! - - ! - - ! - - ! -
-
! - - ! - - ! - - ! - - ! - - ! -
The model was trained on a dataset with information as it was presented on Medium ONLY,
in the posts of Mr. Abraham Loeb who is the sole originator of said information as it was at the time
it was inserted in the 'ds_atlas" dataset used for fine-tuning the base LLM.
The model is offered to its potential users - and should be considered by everyone - as-is, and on a take-it-or-leave-it basis;
that is, given:
-
the inherent hallucinations in any LLM;
-
the performance of the base model;
-
the fine-tuning input/misjudgments/results;
-
the (largely) unpredictable results of the prompt-ing techniques from the part of the user;
-
the (largely) unpredictable results of inference itself,
these shortcomings (either individually, or combined) may result in errors like wrong numerical values, wrong names of people/institutions/etc.,
so it is STRONGLY RECOMMENDED that the model SHOULD NOT be used to extract actionable information
(for example, information that would be used in/for a science paper).
-
! - - ! - - ! - - ! - - ! - - ! -
IF YOU WANT TO USE ACCURATE INFORMATION ABOUT 3I/ATLAS,
YOU ARE HEREBY DIRECTED TO READ THE RESPECTIVE POSTS OF MR. ABRAHAM LOEB ON THIS SUBJECT, WHICH MAKE UP THE SUBSTANCE OF THE ds_atlas DATASET.
- ! - - ! - - ! - - ! - - ! - - ! -
That being said, the users of the model hold any and all responsibility over its use, discussion, and involvement in anything.
SINAPSA Infocomplex will NOT respond to ANY queries related to the model, except to those from Mr. Abraham (Avi) Loeb himself.
- ! - - ! - - ! - - ! - - ! - - ! -
SINAPSA Infocomplex may either update the model, or remove it altogether from public space.
The announcements from our part about updates will ONLY be made on our website at
http://sinapsaro.ro.
Training and evaluation data
Evaluation was performed as follows:
- by inference with LM Studio (https://lmstudio.ai), version 0.3.36 build 1;
- by making comparisons between the output of the model and actual information from the "ds_atlas" dataset.
The values of inference parameters used for evaluation were these:
- Temperature: 0.1
- Top K: 2
- Top P: 0.95
- Min P: 0.9
- Repeat Penalty: 1.14
- optional: Max tokens in reply: 512
Training procedure
The fine-tuning was carried out as follows:
- Supervised Fine-Tuning (SFT);
- by LoRA;
- with LLaMA-Factory 0.9.4 dev (https://github.com/hiyouga/LLaMA-Factory);
- chat template: qwen (as selected in the web UI of LLaMA-Factory 0.9.4 dev);
- on 3,784,704 trainable parameters, out of 467,772,416 from the base model;
- learning_rate: 0.0002 (2e-4)
- on the "ds_atlas" dataset created and curated by us;
- on CPU only (Intel i5);
The new weights are the final checkpoint; no early stopping was performed during fine-tuning.
The merge between the base LLM and the new weights (as .safetensors files) was performed
via a Python application within Microsoft Visual Studio 2026 Community
(
https://devblogs.microsoft.com/visualstudio/visual-studio-2026-is-here-faster-smarter-and-a-hit-with-early-adopters/).
The .safetensors files were converted to GGUF by using llama.cpp b7446 (
https://github.com/ggml-org/llama.cpp) (2025.12).
Framework versions
- Python 12.9
- PEFT 0.15.1
- Transformers 4.51.2
- Pytorch 2.9.1+cpu
- Datasets 3.5.0
- Tokenizers 0.21.1
