GPT-X3-150M-Gundem-LoRA

LoRA adapter for Turkish news headline classification. Requires the pinned base model; this is not a standalone full checkpoint or a general chat assistant.

Base: AxiomicLabs/GPT-X3-150M · revision 541bdc160e37de749378df2cf7de4023c8991fad.

DeepSeek-generated answers were used for assistant-only supervised fine-tuning (response distillation). No DeepSeek weights or logits were used; the student's parameter count is unchanged. The split contains 825 training, 107 validation and 101 test examples. Three epochs, lr=0.0002, LoRA r16/alpha32, accumulation=8, max_length=512, seed=42.

Use

Download this repository, install a suitable CUDA PyTorch build and the supplied requirements, then run from the downloaded folder:

python -m pip install -r requirements.txt
python release_infer.py --model . --headline "İşçiler ücret artışı için greve başladı"

The supplied loader handles architecture compatibility, the exact task template, pinned base revision, CPU fp32 adapter merging and CUDA BF16 inference. Allow CPU RAM for the full fp32 base model and GPU memory for BF16 inference. Loading the base model executes its remote model code. Review it before use. The result contains prediction or an explicit error; generation is not constrained by a JSON grammar.

The model receives only the headline and student-system.txt. Output fields: etiketler (topics), tur (event type), and etki, ciddiyet, aciliyet (integer scores 0–3).

Evaluation

HF/CUDA BF16, greedy, batch=4, maximum 160 new tokens, no JSON grammar. Gündem Score = 30×topic F1 + 20×type accuracy + 30×(1−mean score MAE/3) + 20×exact match (fractional metrics). It measures this task's teacher-label agreement, not general LLM intelligence or an accuracy percentage. Small differences on this previously evaluated, single-seed test do not establish superiority. Batch size and numerical runtime changes can affect outputs. See evaluation.json and release-manifest.json.

Licensing and limitations

Gündem response-distillation LoRA adaptation of AxiomicLabs/GPT-X3-150M, revision 541bdc160e37de749378df2cf7de4023c8991fad. Modified artifact: adapter weights trained on DeepSeek-labelled Turkish headlines. Model and tokenizer licensing follows the upstream license files. Gündem helper code: MIT.

The task uses a small, domain-specific dataset. Inspect predictions before relying on them. Raw headlines, teacher answers and the full dataset are not redistributed. Model licensing is separate from the MIT-licensed Gündem helper code.

Training and benchmark tools · All released models

Mobile ONNX pair

The following standalone full INT8 models run in Gündem Android v0.9.1 or later. These ZIPs do not require downloading the base model or applying an adapter. They use ONNX Runtime for GDN2 inference; they are not GGUF models. tokenizer.gguf inside each ZIP contains only the exact BPE vocabulary, not inference weights.

The original bundle uses the pinned, unadapted checkpoint. The distilled bundle merges the released LoRA into that same checkpoint. INT8 MatMul quantization retains FP32 embeddings and recurrent state. Tokenization matched HF on 106 Turkish/Unicode/whitespace cases. FP32 ONNX logits matched the pinned PyTorch model before quantization (see metadata). Actual production Kotlin/JNI code was exercised on the host JVM, including model installation, constrained JSON, cancellation and reset. This is not a phone-speed measurement. The scores above are from the HF/BF16 experiment, not a new ONNX quality evaluation.

Export code adapted under Apache-2.0 from ahmadw/GPT-X3-150M-ONNX. Model credit and licensing remain with AxiomicLabs. Android app · Export and runtime source.