---
library_name: transformers
tags:
- trl
- sft
- program-of-thought
- reasoning
- physics
license: gpl-3.0
language:
- en
base_model:
- OrionLLM/GRM-2.5
pipeline_tag: text-generation
---
# CoTuGRM-2.5.T2-SFT
Program-of-Thought physics reasoning backbone for numeric educational QA
Model Details ·
Training Data ·
Training Procedure ·
Evaluation ·
Uses
This model is a supervised fine-tune of [`OrionLLM/GRM-2.5`](https://huggingface.co/OrionLLM/GRM-2.5).
The base model has 4B parameters and uses the Qwen3.5 architecture. Team **CoTu** made this model for
**Task 2 — Physics Problems** of the
[EXACT 2026](https://github.com/minhnguyent546/EXACT-2026-CoTu) challenge.
## Model Details
This checkpoint is the Task 2 backbone of the CoTu neuro-symbolic pipeline. The model receives only
the raw text of the physics question. The input has no premises and no formula sheet. The model
finds the applicable physical laws in its parametric memory. Then the model writes a **Python
program** that uses SciPy for the physical constants. A sandboxed interpreter runs the program to
calculate the numeric answer and the unit. For open-ended items, the pipeline runs three parallel
Chain-of-Thought instances at different temperatures. The pipeline then aggregates the three results
into the final JSON output. The model can process the YNU, multiple-choice, and numerical formats.
The covered domains are circuits/electrostatics, mechanics, rotational dynamics, thermodynamics, and
geometric optics.
| | |
|---|---|
| **Developed by** | Team CoTu ([CoTuLabs](https://github.com/cotulabs)) |
| **Model type** | Decoder-only causal LM (Qwen3.5 architecture, `qwen3_5`), fine-tuned with LoRA (merged) |
| **Language** | English |
| **License** | GNU GPL v3.0 |
| **Finetuned from** | [`OrionLLM/GRM-2.5`](https://huggingface.co/OrionLLM/GRM-2.5) |
| **Parameters** | ~4B (within the EXACT 2026 8B open-weight ceiling) |
| **Precision** | BF16 (no quantization) |
## Training Data
We trained the model on **1,452 Task 2 items**. An offline teacher model (**DeepSeek-V4-Pro**) made
new annotations with reasoning traces for these items. The teacher model was not a competitor in the
challenge. Each item has a thinking trace, an answer, an explanation, and a Python program. The
`premises_used` field lists the applied physical laws.
| Stage | Items |
|---|---|
| Official physics set | 1,352 |
| + synthesized MCQs (Qwen3.6-27B) | 100 |
| **Total** | **1,452** |
A second offline teacher model (**Qwen3.6-27B**) made the 100 multiple-choice items from questions
that had only a numerical format. This teacher model was also not a competitor in the challenge.
These new items increase the coverage of the answer formats.
Dataset: [`minhnguyent546/EXACT-2026-CoTu`](https://huggingface.co/datasets/minhnguyent546/EXACT-2026-CoTu)
(Task 2 SFT config: `EXACT2026_dataset_2026-05-15-Physics_Problems_Text_Only-with_synthesized_mcqa_100-val_361-depseek_v4_pro_annotations-v1.0`).
## Training Procedure
We used LoRA supervised fine-tuning. The loss calculation includes only the assistant's reasoning
trace (`...`) and the final response. The prompt tokens are masked. After the
training, we merged the adapters into the base weights. This repository contains the merged
checkpoint.
| Hyperparameter | Value |
|---|---|
| Method | LoRA (rank-stabilized, merged post-training) |
| LoRA rank / alpha / dropout | 16 / 32 / 0.0 |
| Target modules | q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj |
| Learning rate | $1 \times 10^{-4}$ |
| LR schedule / warmup ratio | cosine / 0.03 |
| Batch size / device | 2 |
| Gradient accumulation | 4 |
| Epochs | 1 |
| Weight decay | 0.01 |
| Optimizer | adamw_torch_fused |
| Precision | BF16 |
## Evaluation
The full CoTu pipeline got a **perfect score of 25 of 25 points** for Task 2 in the two automated
rounds. The solver gave a correct answer for each item in all three rounds. Team CoTu got the
**third place overall** and the **highest technical score** in the final round.
## Uses
Use this model with the [CoTu repository](https://github.com/minhnguyent546/EXACT-2026-CoTu)
from version `v0.3.2` (included) to version `v0.5.0` (excluded). The inference suite in these
versions uses the system prompt from training. Versions outside this range can change the system
prompt and reduce model performance.
Use this model for Task 2 physics QA in the CoTu neuro-symbolic pipeline (see the
[instructions here](https://github.com/minhnguyent546/EXACT-2026-CoTu#setup)), together with a sandboxed
Python interpreter. The model can also operate without the pipeline. But the model is trained to
make structured program-of-thought output. This output must go to a downstream interpreter for
execution.
### Sampling parameters
During the challenge, we used the official recommended sampling parameters from
[`Qwen/Qwen3.5-4B`](https://huggingface.co/Qwen/Qwen3.5-4B#using-qwen35-via-the-chat-completions-api):
| Mode | Task type | Sampling parameters |
|---|---|---|
| Thinking | General tasks | `temperature=1.0, top_p=0.95, top_k=20, min_p=0.0, presence_penalty=1.5, repetition_penalty=1.0` |
| Thinking | Precise coding tasks | `temperature=0.6, top_p=0.95, top_k=20, min_p=0.0, presence_penalty=0.0, repetition_penalty=1.0` |
| Non-Thinking (Instruct) | General tasks | `temperature=0.7, top_p=0.8, top_k=20, min_p=0.0, presence_penalty=1.5, repetition_penalty=1.0` |
| Non-Thinking (Instruct) | Reasoning tasks | `temperature=1.0, top_p=1.0, top_k=40, min_p=0.0, presence_penalty=2.0, repetition_penalty=1.0` |
We recommend a maximum output length of 32,768 tokens for most queries. Use 49,152 tokens for
complex tasks, for example math and programming competitions.
## Citation
If you find this work useful in your research, please cite the following paper:
```bibtex
@article{tran2026cotu,
title={CoTu at EXACT 2026: Neuro-Symbolic Reasoning for Transparent Educational QA},
author={Tran, Quoc-Khang and Nguyen, Minh-Thien and Thai, Phu-An and Bui, Xuan-Tung and Ma, Truong-Thanh and Pham, Nguyen-Khang},
journal={arXiv preprint arXiv:2607.14735},
year={2026}
}
```