commercecore-expansion-functional-relation-v1

CommerceCore Expansion — functional_relation adapter v1

A separate project from arghya2030/commercecore-qwen3-1.7b. No shared weights, no shared training run, no shared serving process. That release remains frozen and untouched.

Status: research checkpoint, real partial progress, not a full win. Read the real results below before using this in place of a frontier model.

What this is

A QLoRA rank-16 adapter on an independently-pinned original Qwen/Qwen3-1.7B base (never the Query-merged CommerceCore artifact), fine-tuned to classify the functional relationship between two product listings into substitute, complement, or unrelated.

This is a dedicated, separate adapter from arghya2030/commercecore-expansion-match-v1 (relevance/identity/functional_relation/technical_compatibility as one shared adapter) — trained specifically because a real bug was found in that project's data.

Why this adapter exists: a real bug found after the shared adapter was published

The shared adapter's training data for functional_relation had a development-split bug: every dev row (15/15) was labeled complement, so its reported perfect (1.0) accuracy had only ever been tested on one of three classes. After re-stratifying the split so every class appears in both partitions, the shared adapter's real accuracy is 0.692 (9/13) — a 15.4-point gap behind every tested frontier model (Claude Haiku/Sonnet, GPT-5-mini, GPT-5 all score exactly 0.846 on the same corrected set). This adapter was trained specifically to close that gap.

Real, measured results — genuine improvement, not a win

Data was expanded from 65 to 187 rows (63 distinct item-pair scenarios, up from 15, independently audited by a different model — GPT-5-mini — than the generator — Claude Haiku — same protocol used throughout this project), re-split by label into train (131) / dev (28) / locked_test (28, held out from both training and checkpoint selection).

Training reached a development-set accuracy of 0.893 at its best checkpoint — on dev alone, this looked like a clean win over the 0.846 frontier target. It wasn't, on the honest check:

Best real, honest result: 0.821 — still below the 0.846 frontier target. With only 131 train rows, the training budget used represents dozens of passes over the same small set; the dev set (28 rows, same scenario bank as train) did not fully catch the resulting generalization gap. Checkpoint 450 is published as the model of record for this task — a genuine improvement over the original shared adapter's 0.692 — but does not clear the bar this project requires (beating every tested comparator).

Hugging Face domain-model comparator: NingLab/eCeLLM-S, using its own working instruction style, scores 0.143 on the same locked_test set — beaten by both this adapter and the original shared adapter. Two other comparators (a domain cross-encoder reranker, a general-purpose sentence-embedding model) were checked directly and found unable to separate complement from substitute by any threshold, and are excluded as architecturally not applicable rather than force-scored.

What this is NOT

• Not a win against frontier models — real, disclosed shortfall (0.821 vs. 0.846).
• Not evaluated against the full Hugging Face domain-model comparator set specified in the project's release gate.
• Not load-tested or evaluated under production traffic conditions.

Usage

from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer

base = AutoModelForCausalLM.from_pretrained("Qwen/Qwen3-1.7B")
model = PeftModel.from_pretrained(base, "arghya2030/commercecore-expansion-functional-relation-v1")
tokenizer = AutoTokenizer.from_pretrained("arghya2030/commercecore-expansion-functional-relation-v1")

prompt = (
    "Classify the functional relation: substitute, complement, or unrelated.\n"
    "Espresso Machine - Brew rich, full-bodied espresso shots for your favorite coffee drinks at home.\n"
    "Milk Frother - Create velvety steamed milk and foam for lattes, cappuccinos, and macchiatos.\n"
    "Answer:"
)
inputs = tokenizer(prompt, return_tensors="pt")
out = model.generate(**inputs, max_new_tokens=6, do_sample=False)
print(tokenizer.decode(out[0][inputs["input_ids"].shape[1]:], skip_special_tokens=True))
# complement

Full evaluation methodology and data provenance: github.com/arghya05/commercecore-expansion.