RthItalia/PINDARO-AI-CODE

PINDARO AI CODE

PINDARO AI CODE is the code-specialized release of the Pindaro model family.

Model At A Glance

• Architecture: LlamaForCausalLM
• Model type: llama
• Approx. parameters: ~1.1B
• Precision: float16
• Context length: 2048
• Vocabulary size: 32002
• Languages: English, Italian
• Primary use: code generation and coding assistance

Included Artifacts

Hugging Face format:

• model.safetensors
• config.json
• generation_config.json
• tokenizer.json
• tokenizer.model
• tokenizer_config.json
• special_tokens_map.json
• added_tokens.json

GGUF format:

• pindaro-f16.gguf
• pindaro-q4_k_m.gguf

Release docs:

• release/RELEASE_MANIFEST.json
• release/RELEASE_NOTES.md
• release/SHA256SUMS.txt

Prompt Format

Special tokens:

• <|noesis|> (id 32000)
• <|end|> (id 32001)

Configured chat template uses role sections and appends a code-fence prefix in generation prompt:

{{ bos_token }}{% for message in messages %}<|noesis|>
{% if message['role'] == 'system' %}### System
{{ message['content'] }}
{% elif message['role'] == 'user' %}### Question
{{ message['content'] }}
{% elif message['role'] == 'assistant' %}### Answer
{{ message['content'] }}
{% endif %}<|end|>
{% endfor %}{% if add_generation_prompt %}<|noesis|>
### Answer

{% endif %}

Minimal manual prompt example:

```text
<|noesis|>
### Question
Write a Python function add(a, b).
<|end|>
<|noesis|>
### Answer

## Quickstart (Transformers)

```python
import torch
from transformers import AutoTokenizer, AutoModelForCausalLM

model_id = "RthItalia/PINDARO-AI-CODE"

tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    torch_dtype=torch.float16,
)

messages = [
    {"role": "system", "content": "You are a coding assistant."},
    {"role": "user", "content": "Write a Python function add(a, b)."},
]

inputs = tokenizer.apply_chat_template(
    messages,
    tokenize=True,
    add_generation_prompt=True,
    return_tensors="pt",
)

attention_mask = torch.ones_like(inputs)
outputs = model.generate(
    inputs,
    attention_mask=attention_mask,
    max_new_tokens=120,
    do_sample=False,
)

print(tokenizer.decode(outputs[0], skip_special_tokens=False))

Quickstart (GGUF / llama.cpp)

./llama-cli -m pindaro-q4_k_m.gguf -p "<|noesis|>
### Question
Write a Python function add(a, b).
<|end|>
<|noesis|>
### Answer
```" -n 120

Validation Snapshot

Last internal validation snapshot: 2026-03-02

• HF smoke tests: PASS
• HF mini-eval coding quality: 1.00
• GGUF F16 quality gate: PASS
• GGUF Q4_K_M quality gate: PASS
• Release verdict: publishable: true

Notes:

• Results are from internal sanity checks, not a full public benchmark suite.

Known Limitations

• Generated code can be syntactically correct but logically wrong.
• May emit verbose outputs or repeated scaffolding.
• Always run tests and static checks on generated code.

Safety

• Do not execute generated code in privileged environments without review.
• Use sandboxing for untrusted snippets.
• Add dependency and secret scanning in deployment workflows.

Artifact Checksums (SHA256)

• model.safetensors: f77c27b8babf9fcab83a7dc68ba58934e8c8c031c9f10b4b73e802d4fbfe0cec
• config.json: b37c45060f3e2f5f9b91903c9ccb32f3c21076e809954fda6c01d987cd8f25cc
• generation_config.json: 6ff47e725c0ec6d0f1895670de7ee68e61a4f99703f6c8e89aea6ab14ea02dc3
• tokenizer.json: 51433f06369ac3e597dfa23a811215e3511b8f86588a830ded72344b76a193ee
• tokenizer.model: 9e556afd44213b6bd1be2b850ebbbd98f5481437a8021afaf58ee7fb1818d347
• tokenizer_config.json: a0567c49a117af9af332874cfd333ddd622a09c5e9765131ceee6344cb22a3de
• special_tokens_map.json: d7805e093432afcde852968cdeba3de08a6fe66e77609f4701decb87fc492f33
• added_tokens.json: ece349d292e246eac9a9072c1730f023e61567984a828fb0d25dccb14e3b7592
• pindaro-f16.gguf: bdaaeb6fb712e9a4d952082cf415b05c7d076b33786d39063bbfb3a7e5db2031
• pindaro-q4_k_m.gguf: 5f98cc3454774ed5ed80d71a71adfd0daff760fc9eef0900ddd4f7eda2e20fef