README.md

---
license: apache-2.0
language:
- en
- fr
tags:
- code-generation
- python
- chain-of-thought
- sparse-transformer
- multilingual
- amforge
- sparsemind
library_name: pytorch
pipeline_tag: text-generation
inference: false
model-index:
- name: cofos_v2
  results:
  - task:
      type: text-generation
      name: Python code generation
    metrics:
    - type: real_syntax_valid
      value: 63.0
      name: Real Python syntax validity (held-out)
---

# Cofos v2  Multilingual Python Code Assistant

**Cofos v2** is a 522M-parameter code assistant specialized in Python, with native French/English bilingual support and optional chain-of-thought reasoning. It is the second iteration in the Cofos series by **AMEFORGE**, built on the proprietary **SparseMind** architecture.

This model is designed to produce syntactically correct, executable Python code from natural-language instructions in either French or English, with the ability to emit its reasoning before the code when requested.

---

## Model Summary

| Field | Value |
|---|---|
| **Developer** | AMEFORGE |
| **Architecture** | SparseMind v15 (proprietary) |
| **Parameters** | 522M |
| **Context length** | 2048 tokens |
| **Vocabulary** | 16,384 (custom nexusBPE) |
| **Languages** | French, English |
| **Primary task** | Python code generation |
| **License** | Apache 2.0 |
| **Status** | Active development |

---

## Intended Use

### Primary use cases

- **Python code generation** from natural-language prompts (function specs, class designs, algorithm requests)
- **Bilingual coding assistance** for developers working in French or English
- **Chain-of-thought reasoning** when reasoning steps are useful before the code (toggle via prompt format)
- Integration as a lightweight code assistant in development pipelines where larger models are impractical

### Out-of-scope

This model is **not designed for**:
- General conversation or open-ended dialogue
- Languages other than French and English
- Code in languages other than Python (some JavaScript and Rust tokens are present in the vocabulary but the model has not been trained for general production in those languages)
- Tasks requiring large-context reasoning (>2048 tokens)
- Factual knowledge retrieval, scientific reasoning, or creative writing

Cofos v2 is a specialized coding tool. Use it for what it was built for and pair it with appropriate tools for everything else.

---

## Performance

Evaluated on a held-out set of real Python instruction prompts (no overlap with training data).

| Metric | Value |
|---|---|
| Real-syntax-valid (held-out, n=100) | **63.0%** |
| Validation loss | 3.08 |
| Model size (on disk) | ~2.1 GB (fp32) |

The model has been observed to generate syntactically valid Python with reasonable semantic alignment to short-to-medium instructions. Performance degrades with very long contexts (>1500 tokens) and on instructions that combine multiple distinct subtasks.

## External benchmark

Evaluated on 50 diverse Python prompts (40 EN + 10 FR), comparing against
similarly-sized open models:

| Model | Size | Valid Python | Sem. correct |
|---|---:|---:|---:|
| Cofos v2 (ours) | 522M | 42% | 76%* |
| Qwen2.5-Coder-0.5B | 500M | 84% | 91% |
| SmolLM2-360M | 360M | 92% | 94% |
| Qwen2.5-0.5B | 500M | 74% | 100% |

*Semantic correct = % of test cases passed on the subset of prompts with
verifiable test cases (10 prompts, 30 test cases). Computed only on valid-
syntax generations.

Cofos v2 emits substantially fewer syntactically-valid generations than
comparable open models, but its valid generations show high semantic
accuracy (76%). The gap reflects Cofos v2's much smaller training corpus
(~5000 distilled samples + 16000 real Python instructions vs trillion-token
corpora for the comparison models). Improvement is expected via downstream
fine-tuning (cofos_logo).
---

## Usage

### Loading

```python
from huggingface_hub import hf_hub_download
import torch

# Download checkpoint
checkpoint_path = hf_hub_download(repo_id="AMFORGE/cofos_v2", filename="cofos_model.pt")
tokenizer_path  = hf_hub_download(repo_id="AMFORGE/cofos_v2", filename="cofos_tokenizer.model")

# Loading requires the AMEFORGE inference runtime. Contact AMEFORGE for access
# to the runtime, or use the streaming inference script provided with the model.
```

### Prompt format

Cofos v2 expects a structured prompt format with explicit XML-style tags. The basic pattern is:

```
<instruction>Write a Python function that ...</instruction>
```

For chain-of-thought generation, prefix with a `<thought>` tag:

```
<instruction>Write a Python function that ...</instruction>
<thought>
```

The model will then generate its reasoning, followed by the code block.

---

## Training

Cofos v2 was trained from scratch on a curated mix of:
- Multi-source distilled instruction data with chain-of-thought reasoning (in French and English)
- Real Python instruction-following data from public datasets
- A small synthetic component for algorithmic diversity

Training was conducted with the proprietary SparseMind training pipeline, with periodic safety checkpointing to ensure reproducibility and recovery from interruptions.

**Tokenizer:** [AMFORGE/cofos_tok_v2](https://huggingface.co/AMFORGE/cofos_tok_v2) — a custom SentencePiece model with French-aware coverage, structural XML tags as atomic tokens, and Python keyword/builtin atoms for compact representation of code.

---

## Lineage

```
cofos_tok_v2 (tokenizer)
       ↓
cofos_v2 (this model) — code-specialized from scratch
```

Cofos v2 is a standalone code-specialized model. It is **not** a derivative of any other published model.

---

## Limitations & Biases

- **Capacity**: At 522M parameters, Cofos v2 has limited capacity for complex multi-step reasoning compared to billion-parameter models. Use it for focused coding tasks, not as a general-purpose assistant.
- **Language coverage**: The model is bilingual FR/EN. Prompts in other languages will produce degraded output or fall back to broken English/French.
- **Hallucination**: As with all autoregressive language models, Cofos v2 can produce code that looks plausible but is incorrect. Always test generated code before use.
- **Training data**: While care was taken to use clean, publicly-sourced datasets, the model may reflect biases present in those datasets.
- **No safety alignment**: Cofos v2 has not undergone RLHF or any explicit safety alignment beyond pre-training data curation. It should not be deployed in user-facing products without additional safety layers.

---

## Environmental Considerations

Cofos v2 is intentionally small (522M parameters) to minimize the compute footprint of both training and inference. It can run on a single consumer GPU and is suitable for on-device deployment after appropriate optimization.

---

## License

This model is released under the **Apache 2.0** license. You are free to use, modify, and redistribute it, including for commercial purposes, subject to the terms of the license.

---

## Citation

If you use Cofos v2 in your work, please cite:

```bibtex
@misc{cofos_v2_2026,
  title  = {Cofos v2: A Multilingual Python Code Assistant},
  author = {{AMEFORGE}},
  year   = {2026},
  url    = {https://huggingface.co/AMFORGE/cofos_v2}
}
```

---

## Contact

For questions, collaborations, or access to the AMEFORGE inference runtime:

- **Organization**: AMEFORGE
- **HuggingFace**: [@AMFORGE](https://huggingface.co/AMFORGE)

---

*Cofos is part of a broader family of specialized models being developed by AMEFORGE under the SparseMind architecture program. See the [AMFORGE organization page](https://huggingface.co/AMFORGE) for related work.*