Meridian-AC-Nano
Overview
Meridian-AC-Nano is a compact (~596M parameter) research proofreading and academic-writing model developed by COGERPHERE AI Labs. It is the flagship model of the MERIDIAN 0.1 research programme and is specialized for grammar correction, spelling and punctuation fixes, LaTeX correction, academic tone refinement, and restrained no-over-edit text refinement.
The model is trained on a chat-first, multi-task corpus covering proofreading, LaTeX repair, academic writing, code correction, identity/chat, instruction following, and hard negatives. It loads as a standard causal language model with a ChatML chat template.
Capabilities
Meridian-AC-Nano supports, as verified by the training corpus and per-task evaluation:
- Research proofreading — correcting grammar, spelling, punctuation, and wording while preserving meaning.
- Grammar correction — grammatical and spelling fixes across academic text.
- LaTeX-aware proofreading — fixing syntax and formatting errors in LaTeX source.
- Academic / research writing refinement — tone refinement and text improvement.
- Text refinement / no-over-edit tasks — minimal, meaning-preserving edits (restraint).
Out-of-Scope Use
The model is not intended for production medical, legal, or financial advice; factual knowledge recall; or large-scale general-purpose coding. It is a small model fine-tuned for a specialized proofreading/academic-writing niche.
Evaluation
Results on the held-out test set (n=162, seed 42, greedy decoding) for the flagship run (main_nano):
| Metric | Value |
|---|---|
| ROUGE-1 | 0.9308 |
| ROUGE-2 | 0.9138 |
| ROUGE-L | 0.929 |
| BLEU-4 | 0.8797 |
| Exact match (%) | 74.69 |
| LaTeX validity (%) | 97.53 |
| LaTeX validity (LaTeX subset, %) | 95.24 |
| Quality score | 0.9032 |
Per-Task Results
| Task | n | ROUGE-L | BLEU-4 | Exact match (%) |
|---|---|---|---|---|
| academic | 13 | 0.9744 | 0.9231 | 92.31 |
| chat | 33 | 0.7142 | 0.6064 | 6.06 |
| code | 21 | 1.0000 | 1.0000 | 100.00 |
| hard_negative | 6 | 1.0000 | 1.0000 | 100.00 |
| identity | 4 | 0.7667 | 0.5167 | 50.00 |
| instruction | 10 | 0.9553 | 0.8484 | 70.00 |
| latex | 42 | 0.9978 | 0.9762 | 95.24 |
| proofread | 23 | 0.9884 | 0.9543 | 91.30 |
| restraint | 10 | 1.0000 | 1.0000 | 100.00 |
Training
Fine-tuning was performed on a chat-first multi-task corpus using the Unsloth LoRA pipeline on an NVIDIA RTX 3050 (6 GB Laptop GPU).
| Hyperparameter | Value |
|---|---|
| LoRA rank | 32 |
| LoRA alpha | 64 |
| LoRA dropout | 0.05 |
| Target modules | q_proj, v_proj |
| Epochs | 3 |
| Learning rate | 2e-4 |
| Weight decay | 0.01 |
| Warmup steps | 50 |
| Optimizer | paged_adamw_8bit |
| Batch size / grad accum | 1 / 8 |
| Max sequence length | 2048 |
| Seed | 42 |
Training Data
| Split | Examples |
|---|---|
| Train | 1,275 |
| Validation | 158 |
| Test | 162 |
| Total | 1,595 |
Task categories (all splits): proofread (230), latex (413), academic (130), chat (324), code (198), identity (40), restraint (100), hard_negative (60), instruction (100).
Data sources: manual curation (120), synthetic generation (561), augmented (92), template (822).
Limitations
- Small model: 0.6B parameters; conversational/chat quality is best-effort and lower than larger models.
- Factual knowledge: not a general knowledge or instruction-tuned model at the scale of frontier models.
- Chat exact match is low (6.06%) — expected for open-ended chat responses, which are not verbatim reproducible.
- Domain scope: trained primarily on academic/LaTeX/proofreading content; performance on unrelated domains is not evaluated.
Usage
import torch
from transformers import AutoModelForCausalLM, AutoTokenizer
repo_id = "COGERPHEREAILABS/meridian-ac-nano"
tokenizer = AutoTokenizer.from_pretrained(repo_id)
model = AutoModelForCausalLM.from_pretrained(repo_id, dtype=torch.float16)
messages = [
{
"role": "system",
"content": "You are Meridian, a precise research proofreading assistant. Correct grammar, spelling, punctuation, and wording while preserving the original meaning.",
},
{
"role": "user",
"content": "The results indicates that the model perform very well on the proofreading task.",
},
]
text = tokenizer.apply_chat_template(
messages, tokenize=False, add_generation_prompt=True, return_dict=False
)
inputs = tokenizer(text, return_tensors="pt").to(model.device)
with torch.no_grad():
outputs = model.generate(**inputs, do_sample=False, max_new_tokens=128)
response = tokenizer.decode(outputs[0][inputs["input_ids"].shape[1]:], skip_special_tokens=True)
print(response)
Ollama users can pull the GGUF distribution directly:
ollama run hf.co/COGERPHEREAILABS/meridian-ac-nano-GGUF:Q8_0
License
Apache-2.0. See the base model attribution below for the underlying license terms.
Base Model Attribution
Meridian-AC-Nano is a LoRA fine-tune of Qwen/Qwen3-0.6B (Apache-2.0), sourced via ModelScope (ungated base checkpoint). Full attribution and the base model card are available at https://huggingface.co/Qwen/Qwen3-0.6B. The weights are therefore released under Apache-2.0.
Citation
If you use this model in your research, please cite:
@software{cogerphere_meridian_nano,
author = {Tripathi, Shivanshi and Pandey, Yuvraj},
title = {MERIDIAN 0.1: SLM-Driven Academic Proofreading Models},
year = {2026},
publisher = {COGERPHERE AI Labs},
url = {https://github.com/COGERPHEREAILABS/Meridian-SLM-Driven-Models-Research},
doi = {10.5281/zenodo.21863555}
}
Contact
COGERPHERE AI Labs — GitHub · DOI 10.5281/zenodo.21863555