flint-4bit / validation.json
prince-canuma's picture
Add verified 4bit MLX conversion of RecursiveSI/flint
3064f3a verified
Raw History Blame Contribute Delete
3.87 kB
{
"format": "4bit",
"strict_reload": true,
"finite_prefill_decode_logits": true,
"cache_offsets_correct": true,
"source_sidecars_identical": true,
"native_thinking_template_preserved": true,
"temperature": 0.0,
"max_tokens": 512,
"smoke_tests": [
{
"task": "arithmetic",
"prompt": "What is 17 multiplied by 23? Give the answer.",
"text": "<think>1. The user asks for the product of 17 and 23.\n2. I need to calculate 17 * 23.\n3. Calculation: 17 * 20 = 340, 17 * 3 = 51, sum = 391.\n4. Or, standard multiplication: 17 * 23 = 391.\n5. The user specifically asks to \"Give the answer.\" So I should just provide the number clearly.</think>391",
"answer": "391",
"checks": {
"expected_391": true
},
"generated_tokens": 100,
"seconds": 0.8707173750735819
},
{
"task": "retrieval",
"prompt": "The passcode is 941731. Repeat the passcode exactly.",
"text": "<think>1. The user asks to repeat the passcode exactly, which is \"941731\".\n2. I should just output the string \"941731\" without any additional text, as per the instruction to \"Repeat the passcode exactly.\"\n\nFinal check: The instruction is clear and direct. I will just output the number.</think>941731",
"answer": "941731",
"checks": {
"expected_passcode": true
},
"generated_tokens": 69,
"seconds": 0.59464908298105
},
{
"task": "code",
"prompt": "Write a Python function add(a, b) that returns a + b.",
"text": "<think>We need to write a Python function add(a, b) that returns a + b. This is straightforward. The user likely expects a simple implementation.\n\nWe should provide the code, maybe with a brief explanation. The instruction says \"Write a Python function add(a, b) that returns a + b.\" So we just need to output the function definition.\n\nWe can include type hints or not, but simple is fine.\n\nLet's respond with the code and perhaps a short note.\n\nWe should also consider that the user might be asking for a specific implementation, but it's just basic addition.\n\nI'll write:\n\ndef add(a, b):\n return a + b\n\nMaybe with type hints: def add(a: float, b: float) -> float: but not necessary.\n\nI'll keep it simple.\n\nAlso, we might include a usage example.\n\nBut the instruction is just to write the function, so the answer should be the code.\n\nI'll provide it in a code block.\n\nLet's do that.</think>Here's a simple Python function that returns the sum of two numbers:\n\n```python\ndef add(a, b):\n return a + b\n```\n\nYou can use it like this:\n\n```python\nresult = add(3, 5) # result will be 8\n```",
"answer": "Here's a simple Python function that returns the sum of two numbers:\n\n```python\ndef add(a, b):\n return a + b\n```\n\nYou can use it like this:\n\n```python\nresult = add(3, 5) # result will be 8\n```",
"checks": {
"add_function": true
},
"generated_tokens": 257,
"seconds": 2.1881555418949574
}
],
"all_smoke_checks_pass": true,
"device": {
"device_name": "Apple M5 Max",
"max_recommended_working_set_size": 115448725504,
"memory_size": 137438953472,
"architecture": "applegpu_g17s",
"max_buffer_length": 86586540032,
"resource_limit": 499000
},
"mlx_version": "0.32.4",
"mlx_vlm_version": "0.7.10",
"converter_revision": "6d9ba328abbd6129c7461eda86280aa9fbfd9986",
"python": "3.12.15",
"peak_active_gb": 22.48750338,
"seconds": 4.975497207837179,
"swap_before": "vm.swapusage: total = 10240.00M used = 8744.50M free = 1495.50M (encrypted)",
"swap_after": "vm.swapusage: total = 10240.00M used = 8744.50M free = 1495.50M (encrypted)",
"limits": "Three synthetic smoke prompts and finite/cache checks; not a quality benchmark or proof of parity to BF16/Transformers."
}