Kobarac/gemma4-31b-factual-tool-selector-lora

Gemma 4 31B Factual Tool Selector LoRA

PEFT LoRA adapter for the deterministic-tool routing stage of a factuality judge. For each atomic claim, the selector emits at most one supported tool call or defers to the architecture's fixed claim verifier.

This artifact is the manually selected v1.1 checkpoint 176. It is a selector, not a standalone factuality judge and not an autonomous fact-checking authority.

Lineage

Field	Value
Base model	google/gemma-4-31B-it
Base revision	842da3794eaa0b77d5f08bae87a17459d91ff475
Adapter format	PEFT LoRA
PEFT version	0.20.0
Rank / alpha / dropout	8 / 160 / 0
Target layers	Final eight transformer layers, 52–59
Target modules	q/k/v/o projections and gate/up/down projections
Dataset	Kobarac/gemma4-31b-tool-selector-sft-v1.1
Prompt variant	selector_v1
Maximum calls	1 per claim
Repository commit	ccdf925a5b5e635caccc566367276a311509d73f

Training used a quantized frozen base with trainable BF16 LoRA parameters. The adapter itself is not quantized. A merged BF16 convenience model, when available, must be published separately and verified against adapter-on-base inference.

Evaluation

Checkpoint 176 was selected manually on development data. On the frozen 128-example family evaluation, hybrid correctness improved from 115/128 (89.84%) to 124/128 (96.88%): nine repairs and zero tool-caused regressions. See lineage/selector_checkpoint_decision.json and lineage/manifest.json for the recorded evidence and hashes.

The decision artifact records pending_explicit_activation. Publishing this adapter does not imply that it is enabled by default in the full architecture.

Loading

from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel

base_id = "google/gemma-4-31B-it"
revision = "842da3794eaa0b77d5f08bae87a17459d91ff475"
tokenizer = AutoTokenizer.from_pretrained(base_id, revision=revision)
base = AutoModelForCausalLM.from_pretrained(
    base_id,
    revision=revision,
    torch_dtype="auto",
    device_map="auto",
)
model = PeftModel.from_pretrained(
    base,
    "Kobarac/gemma4-31b-factual-tool-selector-lora",
)

Use the exact selector prompt and deterministic tool registry from the linked dataset or source repository.

Tools and execution contract

The matching dataset publishes the exact executable implementation: source/deterministic_tools.py. The selector can choose exactly one of:

• arithmetic_relation_v1: ratio, percent-change, and relative-percent checks;
• masked_aggregate_v1: masked counts and sums;
• currency_relation_v1: common-basis largest/smallest/total and basis checks;
• scaled_number_v1: coefficient-times-multiplier equality.

Calls use primitive JSON arguments only. Strict validation rejects extra or missing keys, non-finite decimals, oversized vectors, missing currency bases, ambiguous ties, unsupported operations, and partial-claim routes. Rejection, selector deferral, malformed output, and runtime failure all fall back to the fixed neural verifier.

Concrete comparison

On recorded test example E008, a Gemma 4 31B direct judge counted 2 unsupported facts even though all 3 response claims were wrong. Claim extraction exposed all three independently. The fallback verifier still missed the “18% revenue increase” error, but the selector called arithmetic_relation_v1 with Q1 4,940,000, Q2 6,273,800, and claimed value 18; the tool calculated 27 and marked the claim unsupported. The hybrid therefore returned the correct count of 3. See the dataset card for the complete input, call, result, caveat, and links to raw run artifacts.

Integrity

• adapter_config.json: e1f95d84402ed0b49333127db77a51f60db79f12a96798bd1707789193b99002
• adapter_model.safetensors: 29d318f92641297fd1442d698df2b79ecd4a4532defc9296fe375ced19d1621e
• lineage/manifest.json: 1a66f06c8f41d45b2891740c77a5d6298ebe3332d97521c3d42f7db1e3f2a446
• lineage/selector_checkpoint_decision.json: 7252c346e83e5455f624788f986f49f70ea8c6fee0264675486babcf350018b0

Limitations

The selector supports only the versioned primitive registry. Partial-claim routing, malformed plans, and unsupported operations must defer. Performance was measured on a small targeted benchmark and may not generalize to other domains, languages, prompts, or tool implementations.

Licensed under Apache-2.0. Gemma is a trademark of Google LLC; use of the name is descriptive and does not imply endorsement. Preserve upstream notices.