NOVA-1 model card

Field	Value
Model name	NOVA-1
Developer	NOVA AI
Version	1.0
Model type	Decoder-only causal language model, LoRA fine-tune of an instruction-tuned base
Base model	Qwen/Qwen2.5-0.5B-Instruct (© 2024 Alibaba Cloud, Apache 2.0)
Optional base	Qwen/Qwen2.5-1.5B-Instruct for the larger variant
Languages	Hindi (Devanagari), English, Hinglish (romanised Hindi + English)
Parameters	~0.49B base + LoRA adapter (rank 16)
Context length	32,768 tokens for the base; training used 512–768-token sequences
Licence	Apache 2.0 for code and dataset; base weights Apache 2.0; see LICENSE, NOTICE
Repository	this Git repository

Architecture (inherited from the base)

Qwen2.5-0.5B-Instruct is a transformer with 24 layers, grouped-query attention (14 query heads, 2 key/value heads), RoPE positional encoding, SwiGLU MLP, RMSNorm and tied input/output embeddings. NOVA-1 does not change the architecture; it adds trainable low-rank matrices (LoRA) to the linear layers.

Training status — read this first

NOVA-1's training pipeline is complete and verified, but the adapter weights are not produced inside this repository checkout. The environment used to build it has no GPU and no access to huggingface.co, so downloading the 0.5B base model or training it there was impossible. Nothing in this repo pretends otherwise.

To create the actual v1.0 weights:

1. Open notebooks/NOVA1_Finetune_Kaggle_Colab.ipynb on Kaggle or Colab.
2. Enable a free T4 GPU and internet.
3. Run all cells. The notebook trains the adapter, verifies the files, evaluates it against the base model, and writes reports/EVALUATION_REPORT.md.
4. Fill the "Evaluation" and "Training run" sections below with those measured numbers (the notebook prints a template you can copy).

Until that run happens, the only measured numbers in this repository are the pipeline smoke test on a 1.7M random model (reports/smoke/). They are not model results and must never be quoted as NOVA-1 performance.

Intended use

NOVA-1 is intended as a small, inspectable assistant for:

• everyday English/Hindi/Hinglish conversation and study help;
• school-level explanations in science, technology and general knowledge;
• step-by-step arithmetic and simple word problems;
• simple logical reasoning;
• short programming explanations and Python snippets;
• summarising short texts and helping with creative writing;
• education and research into small multilingual fine-tunes.

Out of scope / not recommended

• medical, legal, financial or safety-critical advice;
• anything requiring up-to-date news or internet access (the model has none);
• high-stakes decisions, identity verification or surveillance;
• long-document reasoning, advanced mathematics, competitive programming;
• languages other than Hindi/English beyond what the base model offers;
• generating harmful, illegal or private content.

Training data

A self-authored dataset of 831 records (711 train / 60 validation / 60 test), JSONL, three languages, with grouped splits so translations never leak:

Category	Records
Mathematics	439
Reasoning	102
Science	63
Programming	57
Conversation	31
General knowledge	30
English grammar	21
Technology	20
History	18
Identity	16
Safety / refusal	16
Geography	12
Health	6

Six records are multi-turn conversations (a history field), so the model also learns to answer follow-up questions. The upgrade added 112 curated records: history, geography, English grammar, refusal behaviour and programs in Java, C, C++, Kotlin, SQL, JavaScript, HTML and CSS. | Health | 6 |

Language mix: English 286, Hindi 251, Hinglish 182. Maths answers are computed by Python; code samples were written for this project; no copyrighted books, scraped websites or third-party instruction datasets were used. Full details and known gaps: DATASET_GUIDE.md.

Training procedure

Item	Value
Objective	supervised fine-tuning, next-token prediction
Loss masking	prompt + system tokens masked (-100), loss on assistant answer + EOS
Method	LoRA (optionally QLoRA with 4-bit NF4)
LoRA rank / alpha / dropout	16 / 32 / 0.05
Target modules	q, k, v, o, gate, up, down projections
Epochs	3
Learning rate	2e-4 (0.5B) / 1.5e-4 (1.5B)
Batch × accumulation	4 × 4 on T4 (0.5B); 1 × 16 with 4-bit (1.5B)
Sequence length	512–768
Precision	bf16/fp16 on GPU, fp32 on CPU
Seed	42
Trainable parameters	reported by run_info.json (typically ~1–3% of the model)
Hardware	free Kaggle/Colab T4 (16 GB)

Reproduce with python3 src/nova/train/train_lora.py --config configs/train_lora_0_5b.yaml or the notebook. Exact versions, dataset hashes and timings are stored in run_info.json next to the adapter.

Evaluation

Protocol (implemented in src/nova/eval/evaluate.py):

• validation cross-entropy on assistant tokens and perplexity;
• 50–60 held-out prompts stratified by category and language;
• final-number exact match for numeric items;
• script/language consistency of the reply;
• unigram token-F1 against the reference (rough overlap, not correctness);
• empty-response rate, length, distinct-2, latency;
• raw predictions saved so every claim can be checked.

Results

Real NOVA-1 results: not yet measured. Run the notebook and paste its report here. The report is generated automatically at reports/EVALUATION_REPORT.md.

Pipeline smoke test (not a model result): a 1.7M-parameter randomly initialised Qwen2 model, 40 CPU steps, same code path:

Metric	Base tiny model	After 40 steps
Validation loss	7.6586	7.2681
Perplexity	2118.87	1433.78
Language consistency	0%	0%
Numeric exact match	0%	0%

Random models cannot follow instructions, so quality metrics stay at zero. The numbers only prove that training, saving, reloading and measurement work.

Known evaluation weaknesses

• The test set is small and self-built; it is not MMLU/GSM8K/Indic benchmarks.
• Token-F1 rewards copying words and punishes valid paraphrases.
• Numeric exact match ignores whether the reasoning was correct.
• No human preference study has been done.
• Base-model scores on the same split are required for a fair comparison; the notebook produces both.

Limitations

• Small capacity. 0.5B parameters cannot store broad factual knowledge or perform reliable multi-step reasoning. Expect confident mistakes.
• Maths style bias. 60% of the data is arithmetic, so the model may answer non-maths questions with unnecessary steps.
• Hindi/Hinglish quality depends on the base model and the limited dataset; spelling and dialect variation are not covered.
• No browsing, no tools, no persistent memory. It sees only the current conversation passed by the application.
• Knowledge cutoff is inherited from the base model and is not extended by fine-tuning.
• Prompt sensitivity. Small models drift when the system prompt or chat template changes; keep the NOVA-1 template used in training.
• Safety is not guaranteed. Refusal examples exist, but a 0.5B model can be talked into unsafe or wrong outputs. Add application-level moderation for public deployments.

Bias, safety and privacy

The dataset is small and India-centric in its general-knowledge examples; it does not represent global perspectives. Health answers are general information, not medical advice. No personal data or user conversations were used. The model can reproduce biases present in the base model's pretraining data. For public use, add content filtering, abuse logging (with consent), rate limits and a clear disclaimer that answers can be wrong.

Licence and attribution

• Base model: Qwen2.5-0.5B-Instruct, © 2024 Alibaba Cloud, Apache 2.0.
• Fine-tuning, dataset, code and documentation: © 2026 NOVA AI, Apache 2.0.
• Modifications are listed in NOTICE.
• NOVA AI is not affiliated with or endorsed by Alibaba Cloud or the Qwen team.
• Redistribution must include LICENSE, NOTICE and this model card.

Citation

@misc{nova1_2026,
  title        = {NOVA-1: a small Hindi-English-Hinglish instruction-tuned language model},
  author       = {NOVA AI},
  year         = {2026},
  howpublished = {Fine-tune of Qwen2.5-0.5B-Instruct},
  note         = {Base model by the Qwen team, Alibaba Cloud, Apache 2.0}
}

Contact

Open an issue in this repository for questions, bug reports or dataset corrections.