training::adapter_validator::AdapterValidator¶

Public Functions¶

	Name
	init(self self, float val_loss_threshold =3.0, float behavioral_diff_threshold =5.0, int num_test_prompts =3, int num_inference_prompts =3)
ValidationResult	validate(self self, str niche, str adapter_path, str test_data_path, str base_model_id)

Protected Functions¶

	Name
Optional[bool]	_check_all_zero_weights(Path safetensors_path)
	_decode_tensor(bytes raw, str dtype, list shape, dict dtype_map, np np)
bool	_check_all_zero_npz(Path npz_path)
float	_compute_validation_loss(self self, Path test_data_path)
float	_compute_behavioral_diff(self self, Path test_data_path, str base_model_id, Path adapter_dir)

Protected Attributes¶

	Name
	_val_loss_threshold
	_behavioral_diff_threshold
	_num_test_prompts
	_num_inference_prompts

Detailed Description¶

class training::adapter_validator::AdapterValidator;

Multi-prong post-training adapter quality validator.

Checks: (a) loadability — adapter files exist, load cleanly, have non-zero norms,
(b) validation loss — below per-specialist threshold,
(c) behavioral diff — inference output differs from base model.

Public Functions Documentation¶

function init¶

__init__(
    self self,
    float val_loss_threshold =3.0,
    float behavioral_diff_threshold =5.0,
    int num_test_prompts =3,
    int num_inference_prompts =3
)

function validate¶

ValidationResult validate(
    self self,
    str niche,
    str adapter_path,
    str test_data_path,
    str base_model_id
)

Run full multi-prong validation.```


## Protected Functions Documentation

### function _check_all_zero_weights

```python
static Optional[bool] _check_all_zero_weights(
    Path safetensors_path
)

Check if all tensors in a safetensors file have near-zero norms.

Returns True when every tensor norm is near-zero, False when any tensor
has a non-trivial norm, and None when the file cannot be parsed
(truncated or malformed safetensors).

function _decode_tensor¶

static _decode_tensor(
    bytes raw,
    str dtype,
    list shape,
    dict dtype_map,
    np np
)

Decode raw safetensors bytes into a numeric numpy array.

Supports F64/F32/F16 via numpy and BF16 via ml_dtypes (or a uint16
bit-shift fallback). Returns None for unsupported dtypes.

function _check_all_zero_npz¶

static bool _check_all_zero_npz(
    Path npz_path
)

Check if all tensors in an npz file have near-zero norms.```


### function _compute_validation_loss

```python
float _compute_validation_loss(
    self self,
    Path test_data_path
)

Compute mean cross-entropy loss on test data.```


### function _compute_behavioral_diff

```python
float _compute_behavioral_diff(
    self self,
    Path test_data_path,
    str base_model_id,
    Path adapter_dir
)

Compute output divergence between base model and adapter.

Real base-vs-adapter inference requires MLX model loading, which is
deferred to a later phase. Until then this check fails closed: it
returns 0.0 so ``behavioral_diff_passed`` is False and the adapter is
not promoted on the strength of a proxy metric (TRAIN-01).

Protected Attributes Documentation¶

variable _val_loss_threshold¶

_val_loss_threshold;

variable _behavioral_diff_threshold¶

_behavioral_diff_threshold;

variable _num_test_prompts¶

_num_test_prompts;

variable _num_inference_prompts¶

_num_inference_prompts;

Updated on 2026-10-09 at 03:59:43 +0000