Instructions to use veriloop-lab/veriloop-coder-e1 with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use veriloop-lab/veriloop-coder-e1 with Transformers:

# Use a pipeline as a high-level helper
from transformers import pipeline

pipe = pipeline("text-generation", model="veriloop-lab/veriloop-coder-e1")
messages = [
    {
        "role": "user",
        "content": [
            {"type": "image", "url": "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/p-blog/candy.JPG"},
            {"type": "text", "text": "What animal is on the candy?"}
        ]
    },
]
pipe(text=messages)

# Load model directly
from transformers import AutoProcessor, AutoModelForImageTextToText

processor = AutoProcessor.from_pretrained("veriloop-lab/veriloop-coder-e1")
model = AutoModelForImageTextToText.from_pretrained("veriloop-lab/veriloop-coder-e1")
messages = [
    {
        "role": "user",
        "content": [
            {"type": "image", "url": "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/p-blog/candy.JPG"},
            {"type": "text", "text": "What animal is on the candy?"}
        ]
    },
]
inputs = processor.apply_chat_template(
	messages,
	add_generation_prompt=True,
	tokenize=True,
	return_dict=True,
	return_tensors="pt",
).to(model.device)

outputs = model.generate(**inputs, max_new_tokens=40)
print(processor.decode(outputs[0][inputs["input_ids"].shape[-1]:]))

PEFT
How to use veriloop-lab/veriloop-coder-e1 with PEFT:
```
Task type is invalid.
```
Notebooks
Google Colab
Kaggle
Local Apps

vLLM

How to use veriloop-lab/veriloop-coder-e1 with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "veriloop-lab/veriloop-coder-e1"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "veriloop-lab/veriloop-coder-e1",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/veriloop-lab/veriloop-coder-e1

SGLang

How to use veriloop-lab/veriloop-coder-e1 with SGLang:

Install from pip and serve model

# Install SGLang from pip:
pip install sglang
# Start the SGLang server:
python3 -m sglang.launch_server \
    --model-path "veriloop-lab/veriloop-coder-e1" \
    --host 0.0.0.0 \
    --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "veriloop-lab/veriloop-coder-e1",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker images

docker run --gpus all \
    --shm-size 32g \
    -p 30000:30000 \
    -v ~/.cache/huggingface:/root/.cache/huggingface \
    --env "HF_TOKEN=<secret>" \
    --ipc=host \
    lmsysorg/sglang:latest \
    python3 -m sglang.launch_server \
        --model-path "veriloop-lab/veriloop-coder-e1" \
        --host 0.0.0.0 \
        --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "veriloop-lab/veriloop-coder-e1",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Docker Model Runner
How to use veriloop-lab/veriloop-coder-e1 with Docker Model Runner:
```
docker model run hf.co/veriloop-lab/veriloop-coder-e1
```

veriloop-coder-e1 / uncertainty_adapter_train_result.json

ConorWang

Upload uncertainty adapter artifacts to repo root

a856587 verified about 1 month ago

raw

history blame

21 kB

	{
	"artifacts": {
	"adapter_dir": "/private/wang_libo/veriloop_coder_e1/outputs/uncertainty_qwen36_rootfix_run1/adapter",
	"best_checkpoint_manifest": "/private/wang_libo/veriloop_coder_e1/outputs/uncertainty_qwen36_rootfix_run1/best_checkpoint_manifest.json",
	"epoch_history": "/private/wang_libo/veriloop_coder_e1/outputs/uncertainty_qwen36_rootfix_run1/epoch_history.json",
	"eval_jsonl": "/private/wang_libo/veriloop_coder_e1/outputs/uncertainty_qwen36_rootfix_run1/uncertainty_eval.jsonl",
	"host_manifest": "/private/wang_libo/veriloop_coder_e1/outputs/uncertainty_qwen36_rootfix_run1/host_manifest.json",
	"plan_json": "/private/wang_libo/veriloop_coder_e1/outputs/uncertainty_qwen36_rootfix_run1/uncertainty_adapter_plan.json",
	"tokenizer_dir": "/private/wang_libo/veriloop_coder_e1/outputs/uncertainty_qwen36_rootfix_run1/tokenizer",
	"train_jsonl": "/private/wang_libo/veriloop_coder_e1/outputs/uncertainty_qwen36_rootfix_run1/uncertainty_train.jsonl",
	"training_manifest": "/private/wang_libo/veriloop_coder_e1/outputs/uncertainty_qwen36_rootfix_run1/uncertainty_training_manifest.json",
	"uncertainty_head": "/private/wang_libo/veriloop_coder_e1/outputs/uncertainty_qwen36_rootfix_run1/uncertainty_head.pt"
	},
	"dataset": {
	"eval_modes": {
	"conflicting_evidence": 5,
	"evidence_gap": 5,
	"exec_required": 5,
	"high_risk": 5,
	"low_uncertainty": 5,
	"patch_pending": 5,
	"reverse_engineering_ambiguity": 5,
	"self_check_failure": 5,
	"spec_mismatch": 5,
	"validator_negation": 5,
	"worktree_conflict": 5
	},
	"eval_size": 55,
	"train_modes": {
	"conflicting_evidence": 14,
	"evidence_gap": 14,
	"exec_required": 14,
	"high_risk": 14,
	"low_uncertainty": 14,
	"patch_pending": 14,
	"reverse_engineering_ambiguity": 14,
	"self_check_failure": 14,
	"spec_mismatch": 14,
	"validator_negation": 14,
	"worktree_conflict": 14
	},
	"train_size": 154
	},
	"eval_metrics": {
	"adapter_exported": true,
	"auto_lora_from_ia3": false,
	"best_epoch": 2,
	"best_quality_score": 0.8625935807221907,
	"count": 55,
	"mae": {
	"u_answer": 0.15174226462841034,
	"u_evidence": 0.19610758125782013,
	"u_exec": 0.18561214208602905,
	"u_risk": 0.1553734689950943,
	"u_spec": 0.21633382141590118
	},
	"mean_mae": 0.18103384971618652,
	"mean_rmse": 0.24169571697711945,
	"moderate_accuracy": 0.6727272727272727,
	"peft_method": "lora_narrow",
	"quality_score": 0.8625935807221907,
	"rmse": {
	"u_answer": 0.18724055588245392,
	"u_evidence": 0.22527915239334106,
	"u_exec": 0.25238174200057983,
	"u_risk": 0.20667441189289093,
	"u_spec": 0.3369026482105255
	},
	"tight_accuracy": 0.4,
	"used_peft": true,
	"weighted_mae": 0.18083095811830807,
	"weighted_rmse": 0.24125460771003793
	},
	"plan": {
	"adapter_alpha": 16,
	"adapter_dropout": 0.05,
	"adapter_rank": 8,
	"backbone_manifest": {
	"host_inventory_size": 79
	},
	"config": {
	"allow_backbone_bridge": false,
	"backbone": "/public/wang_libo/veriloop_coder_e1/model",
	"bf16": true,
	"cache_dir": null,
	"cpu_max_memory_gib": 96,
	"dataset_jsonl": null,
	"dim_weights": {
	"u_answer": 1.1,
	"u_evidence": 1.35,
	"u_exec": 1.35,
	"u_risk": 1.55,
	"u_spec": 1.25
	},
	"early_stopping_min_delta": 0.001,
	"early_stopping_patience": 2,
	"enable_synthetic_dataset": true,
	"eval_jsonl": null,
	"eval_samples_per_mode": 5,
	"fp16": false,
	"gpu_max_memory_gib": 44,
	"gradient_accumulation_steps": 16,
	"host_dropout": 0.0,
	"learning_rate": 2e-05,
	"local_files_only": true,
	"logging_steps": 10,
	"max_grad_norm": 1.0,
	"max_length": 1664,
	"min_epochs_before_early_stop": 2,
	"num_train_epochs": 4.0,
	"output_dir": "./outputs/uncertainty_qwen36_rootfix_run1",
	"per_device_eval_batch_size": 1,
	"per_device_train_batch_size": 1,
	"prefer_best_checkpoint_export": true,
	"probe_dropout": 0.03,
	"probe_hidden_factor": 0.75,
	"product_line": "veriloop_coder",
	"quantization_mode": "4bit",
	"quantization_required": false,
	"require_trainable_targets": true,
	"revision": null,
	"run_post_train_eval": true,
	"save_best_checkpoint": true,
	"seed": 11,
	"selection_mode": "minimal",
	"train_samples_per_mode": 14,
	"training_mode": "mounted_head",
	"trust_remote_code": true,
	"use_double_quant": true,
	"warmup_ratio": 0.05,
	"weight_decay": 0.0,
	"weighted_mae_penalty": 0.5,
	"weighted_rmse_penalty": 0.5
	},
	"dataset_summary": {
	"eval_modes": {
	"conflicting_evidence": 5,
	"evidence_gap": 5,
	"exec_required": 5,
	"high_risk": 5,
	"low_uncertainty": 5,
	"patch_pending": 5,
	"reverse_engineering_ambiguity": 5,
	"self_check_failure": 5,
	"spec_mismatch": 5,
	"validator_negation": 5,
	"worktree_conflict": 5
	},
	"eval_size": 55,
	"train_modes": {
	"conflicting_evidence": 14,
	"evidence_gap": 14,
	"exec_required": 14,
	"high_risk": 14,
	"low_uncertainty": 14,
	"patch_pending": 14,
	"reverse_engineering_ambiguity": 14,
	"self_check_failure": 14,
	"spec_mismatch": 14,
	"validator_negation": 14,
	"worktree_conflict": 14
	},
	"train_size": 154
	},
	"head_strategy": "host_head",
	"notes": [
	"Primary route is host-surface-first uncertainty training.",
	"Validator receipts, self-check failure, reverse-engineering ambiguity, worktree conflicts, and patch continuity are first-class signals.",
	"DualPath, Full AttnRes, mHC, routers, experts, and broad attention-layer PEFT stay excluded.",
	"This adapter should improve runtime uncertainty routing, not general coding free-formity."
	],
	"peft_method": "lora_narrow",
	"recipe": {
	"adapter_family": "uncertainty",
	"backbone": "/public/wang_libo/veriloop_coder_e1/model",
	"backbone_family": "qwen_dense",
	"excluded_patterns": [
	"(?i)\\bdualpath\\b",
	"(?i)\\bmhc\\b",
	"(?i)\\bfull[_\\- ]?attnres\\b",
	"(?i)\\battnres(_full)?\\b",
	"(?i)\\brouter\\b",
	"(?i)\\bexperts?\\b",
	"(?i)\\bmoe\\b.*\\b(gate\|router\|expert)\\b",
	"(?i)\\brope\\b.*\\b(freq\|inv_freq\|theta\|rotary)\\b",
	"(?i)\\bkvcache\\b",
	"(?i)\\bposition_embedding\\b",
	"(?i)\\bembed(tokens\|ding)?\\b",
	"(?i)\\blm_head\\b"
	],
	"harness_constraints": [
	"Harness Engineering remains the primary convergence layer.",
	"Adapter must not bypass runtime orchestrator / validator / rollback loops.",
	"Adapter outputs remain subordinate to VeriLoop control-plane decisions.",
	"Adapter must not create hidden prompt-style memory authority.",
	"Adapter must support bounded uncertainty calibration rather than generic hesitation.",
	"Validator and receipt evidence must remain able to update uncertainty."
	],
	"hyperparams": {
	"alpha": 16,
	"bias": "none",
	"dropout": 0.05,
	"fan_in_fan_out": false,
	"modules_to_save": [
	"input_layernorm"
	],
	"r": 8,
	"task_type": "CAUSAL_LM"
	},
	"merge_policy": "merge_after_guard",
	"metadata": {
	"allow_backbone_bridge": false,
	"allow_vla_action_expert": false,
	"harness_first": true,
	"prefer_explicit_heads": true,
	"prefer_qlora_for_backbone_bridge": true,
	"require_harness_first": true,
	"selector_group_count": 2,
	"strict_narrow_scope": true,
	"trainer": "veriloop.uncertainty_adapter_trainer.v5.qwen36",
	"uncertainty_training": true
	},
	"notes": [
	"Backbone bridge tuning disabled explicitly; selector stays on custom surfaces or no-op.",
	"Backbone family inferred as qwen_dense.",
	"PEFT method resolved as lora_narrow.",
	"Recipe is harness-first: runtime convergence remains in VeriLoop control-plane + harness, not in broad weight surgery.",
	"Block AttnRes, DualPath, mHC hooks, RoPE, KV-cache, and broad MoE routing remain structurally excluded."
	],
	"peft_method": "lora_narrow",
	"precision_policy": "auto",
	"product_line": "veriloop_coder",
	"regression_requirements": [
	"Must pass PEFT regression guard structural policy checks.",
	"Must not introduce forbidden backbone/serving structural targets.",
	"Must preserve harness regression envelope for the selected product line.",
	"Budgeted uncertainty convergence must not regress.",
	"Uncertainty calibration must not collapse into generic caution."
	],
	"target_groups": [
	{
	"alpha": 16,
	"dropout": 0.0,
	"name": "group_1_custom_control_head",
	"rank": 8,
	"rationale": "Prefer explicit uncertainty / calibration heads over backbone surgery.",
	"surface": "custom_control_head",
	"target_modules": [
	"uncertainty_head",
	"uncertainty_head.calibration_mlp",
	"uncertainty_head.proj"
	]
	},
	{
	"alpha": 16,
	"dropout": 0.0,
	"name": "group_2_custom_validator_bridge",
	"rank": 8,
	"rationale": "Validation and rollback fidelity should prefer explicit validator / rollback bridges.",
	"surface": "custom_validator_bridge",
	"target_modules": [
	"failure_signal_bridge",
	"failure_signal_bridge.rollback_bridge",
	"rollback_adapter",
	"rollback_adapter.head",
	"sandbox_rollback_bridge",
	"sandbox_rollback_bridge.adapter",
	"validator_feedback_bridge",
	"validator_feedback_bridge.adapter",
	"validator_feedback_loop.rollback_adapter"
	]
	}
	],
	"target_modules": [
	"uncertainty_head",
	"uncertainty_head.calibration_mlp",
	"uncertainty_head.proj",
	"failure_signal_bridge",
	"failure_signal_bridge.rollback_bridge",
	"rollback_adapter",
	"rollback_adapter.head",
	"sandbox_rollback_bridge",
	"sandbox_rollback_bridge.adapter",
	"validator_feedback_bridge",
	"validator_feedback_bridge.adapter",
	"validator_feedback_loop.rollback_adapter"
	],
	"version": "veriloop.lora_recipe_veriloop.v2"
	},
	"selected_target_modules": [
	"uncertainty_head",
	"uncertainty_head.calibration_mlp",
	"uncertainty_head.proj"
	],
	"target_selection": {
	"backbone_archetype": "qwen_dense",
	"exclusions": [
	{
	"pattern": "(^\|\\.)lm_head($\|\\.)",
	"reason": "Do not retune final token head; too broad and evaluation-heavy."
	},
	{
	"pattern": "(^\|\\.)embed_tokens($\|\\.)",
	"reason": "Embedding surgery risks broad semantic drift."
	},
	{
	"pattern": "(^\|\\.)norm($\|\\.)",
	"reason": "Global norm tuning can destabilize calibration across scenes."
	},
	{
	"pattern": "attnres\|attention_residual",
	"reason": "Block AttnRes may be mounted structurally but is never a PEFT target."
	},
	{
	"pattern": "dualpath",
	"reason": "DualPath is serving/runtime infrastructure only."
	},
	{
	"pattern": "mhc\|hyper[-_]?connection",
	"reason": "mHC-inspired stability hooks remain structural, not PEFT surfaces."
	},
	{
	"pattern": "rope\|rotary",
	"reason": "RoPE/context surgery is handled architecturally, not by narrow PEFT here."
	},
	{
	"pattern": "kvcache\|kv_cache",
	"reason": "KV-cache runtime surfaces are not PEFT targets."
	},
	{
	"pattern": "(^\|\\.)memory(_store\|_bank)?($\|\\.)",
	"reason": "Persistent memory stores are harness/runtime policy surfaces, not PEFT targets."
	}
	],
	"inventory_size": 79,
	"inventory_source": "provided_names",
	"notes": [
	"Harness Engineering is primary; PEFT is limited to obedience-facing, interface-facing support surfaces.",
	"Backbone bridge tuning disabled explicitly; selector stays on custom surfaces or no-op."
	],
	"recommended_training": "ia3_head_only",
	"request": {
	"allow_backbone_bridge": false,
	"allow_vla_action_expert": false,
	"backbone": "/public/wang_libo/veriloop_coder_e1/model",
	"explicit_exclude_patterns": [],
	"explicit_include_patterns": [],
	"intents": [
	"uncertainty",
	"validator_alignment",
	"harness_alignment",
	"runtime_protocol",
	"session_continuity",
	"worktree_discipline"
	],
	"prefer_qlora_for_backbone_bridge": true,
	"product_line": "veriloop_coder",
	"selection_mode": "minimal"
	},
	"selected_groups": [
	{
	"alpha": 16,
	"dropout": 0.0,
	"exclude_patterns": [
	"(^\|\\.)lm_head($\|\\.)",
	"(^\|\\.)embed_tokens($\|\\.)",
	"(^\|\\.)norm($\|\\.)",
	"attnres\|attention_residual",
	"dualpath",
	"mhc\|hyper[-_]?connection",
	"rope\|rotary",
	"kvcache\|kv_cache",
	"(^\|\\.)memory(_store\|_bank)?($\|\\.)"
	],
	"include_patterns": [
	"(^\|\\.)(uncertainty_head\|uncertainty_adapter\|calib(_head\|ration_adapter)?)($\|\\.)"
	],
	"intents": [
	"uncertainty"
	],
	"layer_window": {
	"mode": "all",
	"value": 0.0
	},
	"matched_module_names": [
	"uncertainty_head",
	"uncertainty_head.calibration_mlp",
	"uncertainty_head.proj"
	],
	"name": "group_1_custom_control_head",
	"rank": 8,
	"rationale": "Prefer explicit uncertainty / calibration heads over backbone surgery.",
	"risk": "low",
	"surface": "custom_control_head"
	},
	{
	"alpha": 16,
	"dropout": 0.05,
	"exclude_patterns": [
	"(^\|\\.)lm_head($\|\\.)",
	"(^\|\\.)embed_tokens($\|\\.)",
	"(^\|\\.)norm($\|\\.)",
	"attnres\|attention_residual",
	"dualpath",
	"mhc\|hyper[-_]?connection",
	"rope\|rotary",
	"kvcache\|kv_cache",
	"(^\|\\.)memory(_store\|_bank)?($\|\\.)"
	],
	"include_patterns": [
	"(^\|\\.)(runtime_harness\|query_runtime\|task_brief\|task_runtime\|tool_protocol\|permission_context\|worktree\|session_state\|request_normalizer\|action_allowlist\|constraint_guard\|progress_state\|workspace_snapshot\|repo_contract\|knowledge_entry\|completion_criteria\|search_bridge\|sandbox_search_bridge)(_adapter\|_bridge\|_head)?($\|\\.)",
	"(^\|\\.)(toolspec\|tool(_call)?(_grammar\|_interface)?\|harness\|validator\|rollback\|receipt\|patch\|permission\|session\|worktree)(_adapter\|_bridge\|_head)?($\|\\.)"
	],
	"intents": [
	"harness_alignment",
	"runtime_protocol",
	"session_continuity",
	"worktree_discipline"
	],
	"layer_window": {
	"mode": "all",
	"value": 0.0
	},
	"matched_module_names": [
	"failure_signal_bridge.rollback_bridge",
	"request_normalizer",
	"request_normalizer.adapter",
	"rollback_adapter",
	"rollback_adapter.head",
	"runtime_harness_adapter",
	"runtime_harness_adapter.bridge",
	"tool_protocol_adapter",
	"tool_protocol_adapter.bridge",
	"toolspec_bridge",
	"toolspec_bridge.adapter",
	"toolspec_head",
	"toolspec_head.param_schema_adapter",
	"toolspec_head.postcondition_adapter",
	"toolspec_head.precondition_adapter",
	"toolspec_head.receipt_formatter",
	"toolspec_head.trigger_gate",
	"toolspec_head.validator_gate",
	"validator_feedback_loop.rollback_adapter"
	],
	"name": "group_2_custom_runtime_harness_bridge",
	"rank": 8,
	"rationale": "Runtime / harness obedience should attach to explicit interface bridges before any backbone fallback.",
	"risk": "low",
	"surface": "custom_runtime_harness_bridge"
	},
	{
	"alpha": 16,
	"dropout": 0.0,
	"exclude_patterns": [
	"(^\|\\.)lm_head($\|\\.)",
	"(^\|\\.)embed_tokens($\|\\.)",
	"(^\|\\.)norm($\|\\.)",
	"attnres\|attention_residual",
	"dualpath",
	"mhc\|hyper[-_]?connection",
	"rope\|rotary",
	"kvcache\|kv_cache",
	"(^\|\\.)memory(_store\|_bank)?($\|\\.)"
	],
	"include_patterns": [
	"(^\|\\.)(validator_feedback\|sandbox_result_validator\|sandbox_rollback_bridge\|failure_signal\|rollback)(_adapter\|_bridge\|_head)?($\|\\.)"
	],
	"intents": [
	"validator_alignment"
	],
	"layer_window": {
	"mode": "all",
	"value": 0.0
	},
	"matched_module_names": [
	"failure_signal_bridge",
	"failure_signal_bridge.rollback_bridge",
	"rollback_adapter",
	"rollback_adapter.head",
	"sandbox_rollback_bridge",
	"sandbox_rollback_bridge.adapter",
	"validator_feedback_bridge",
	"validator_feedback_bridge.adapter",
	"validator_feedback_loop.rollback_adapter"
	],
	"name": "group_3_custom_validator_bridge",
	"rank": 8,
	"rationale": "Validation and rollback fidelity should prefer explicit validator / rollback bridges.",
	"risk": "low",
	"surface": "custom_validator_bridge"
	},
	{
	"alpha": 8,
	"dropout": 0.0,
	"exclude_patterns": [
	"(^\|\\.)lm_head($\|\\.)",
	"(^\|\\.)embed_tokens($\|\\.)",
	"(^\|\\.)norm($\|\\.)",
	"attnres\|attention_residual",
	"dualpath",
	"mhc\|hyper[-_]?connection",
	"rope\|rotary",
	"kvcache\|kv_cache",
	"(^\|\\.)memory(_store\|_bank)?($\|\\.)"
	],
	"include_patterns": [
	"(^\|\\.)(memory_boundary_guard\|episodic_memory\|session_compactor)(_adapter\|_bridge\|_head)?($\|\\.)"
	],
	"intents": [
	"session_continuity"
	],
	"layer_window": {
	"mode": "all",
	"value": 0.0
	},
	"matched_module_names": [
	"episodic_memory",
	"episodic_memory.adapter",
	"memory_boundary_guard",
	"memory_boundary_guard.adapter",
	"memory_boundary_guard.rollback_filter",
	"session_compactor",
	"session_compactor.adapter"
	],
	"name": "group_4_custom_memory_boundary_bridge",
	"rank": 4,
	"rationale": "Session continuity should bind to boundary-aware memory packet surfaces rather than broad backbone tuning.",
	"risk": "low",
	"surface": "custom_memory_boundary_bridge"
	}
	],
	"upstream_profile_hint": null,
	"version": "veriloop.peft_target_selector.v2",
	"warnings": []
	},
	"training_mode": "mounted_head",
	"version": "veriloop.uncertainty_adapter_trainer.v5.qwen36",
	"warnings": [
	"Harness Engineering is primary; PEFT is limited to obedience-facing, interface-facing support surfaces.",
	"Backbone bridge tuning disabled explicitly; selector stays on custom surfaces or no-op."
	]
	},
	"status": "trained",
	"train_metrics": {
	"adapter_exported": true,
	"auto_lora_from_ia3": false,
	"best_epoch": 2,
	"best_quality_score": 0.8625935807221907,
	"epochs_completed": 4,
	"loss": 0.009006613283418119,
	"micro_batches": 154,
	"micro_batches_total": 616,
	"optimizer_steps": 10,
	"optimizer_steps_total": 40,
	"peft_method": "lora_narrow",
	"used_peft": true
	},
	"version": "veriloop.uncertainty_adapter_trainer.v5.qwen36",
	"warnings": [
	"Harness Engineering is primary; PEFT is limited to obedience-facing, interface-facing support surfaces.",
	"Backbone bridge tuning disabled explicitly; selector stays on custom surfaces or no-op.",
	"Synthetic holdout split used for quality-oriented train/eval separation."
	]
	}