A LoRA adapter that makes Qwen/Qwen2.5-1.5B-Instruct reliably convert free-text descriptions of small-business processes into a strict 10-field structured workflow analysis.

Trained on ArashEslam/business-workflow-extraction.

Results

Evaluated on 40 held-out records (28 synthetic + 12 real human-written). The meaningful comparison is against a 2-shot baseline, not zero-shot: a zero-shot 1.5B fails mostly on JSON formatting, so a naive before/after would report a formatting win as a reasoning win.

Metric	Base (0-shot)	Base (2-shot)	This adapter
Schema-valid rate	70%	80%	98%
AI-step grounding	0.230	0.326	0.420
Length ratio vs reference	0.389	0.702	0.838
LLM-judge rubric (1-5)	3.04	3.23	3.41

Blinded pairwise vs the 2-shot baseline, judged in both orders to control position bias: 30 wins / 7 ties / 3 losses (91% excluding ties), margin 95% CI [+0.47, +0.85], position inconsistency 18%.

Strict-JSON compliance rises from 0% to ~98%, but this is deliberately deprioritised: all base generations were markdown-fenced, and an identical repair function applied to every arm recovers 100% of them. That swing is formatting discipline, not workflow quality.

Training


Base model	`Qwen/Qwen2.5-1.5B-Instruct`
Method	LoRA (bf16; 4-bit QLoRA path available)
Rank / alpha / dropout	16 / 32 / 0.05
Target modules	`q_proj`, `k_proj`, `v_proj`, `o_proj`, `gate_proj`, `up_proj`, `down_proj`
Trainable params	18,464,768 (1.18%)
LR / schedule	0.0002 / cosine, warmup 0.03
Epochs / optimizer steps	3 / 90
Effective batch	1 × 4 grad-accum
Max sequence length	2560
Loss	completion tokens only (prompt masked to -100)

Dev loss by epoch:

Epoch	Dev loss
1	1.1796
2	1.1320
3	1.1375

Dev loss plateaus after epoch 2 while train loss keeps falling. The best checkpoint (epoch 2) is what ships here. The epoch-3 uptick is far too small relative to the 12-record dev set to call overfitting — the honest reading is that the model extracted what it can from 120 examples.

LoRA targets MLP as well as attention projections: the measured baseline deficit was content generation (empty reasoning fields, ~1/3 of reference length), not attention routing.

Usage

python

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel

BASE = "Qwen/Qwen2.5-1.5B-Instruct"
ADAPTER = "ArashEslam/qwen2.5-1.5b-workflow-lora"

tok = AutoTokenizer.from_pretrained(BASE)
model = AutoModelForCausalLM.from_pretrained(BASE, torch_dtype=torch.bfloat16).to("cuda")
model = PeftModel.from_pretrained(model, ADAPTER)
model.eval()

messages = [
    {"role": "system", "content": "You are an operations analyst. You read descriptions of how a small business currently runs a process and turn them into a structured workflow analysis."},
    {"role": "user", "content": "<the instructions + business process description>"},
]
text = tok.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tok(text, return_tensors="pt", add_special_tokens=False).to(model.device)
out = model.generate(**inputs, max_new_tokens=2048, do_sample=False)
print(tok.decode(out[0][inputs.input_ids.shape[1]:], skip_special_tokens=True))

The exact prompt template matters — the adapter was trained against one fixed template. Use src/nextverse/prompts/task.py from the companion repository to build prompts, rather than improvising the instruction text.

Limitations

Trained on synthetic references. The targets encode one frontier model's idea of a good workflow analysis, not a validated business standard.
Small scale. 120 training examples, 40 evaluation records. Per-slice findings are indicative, not precise.
Two fields regressed versus the baseline: bottlenecks_and_risks (3.19 → 2.77) and human_approvals_controls (2.53 → 2.35). Not under-generation — the model learned the reference's rhetorical form (long causal statements) without its evidential specificity (named entities and figures drawn from the input). Style is cheap to learn from 120 examples; specificity is not.
JSON compliance is real but shallow. One evaluation record produced Python-style string concatenation inside JSON when the input stated a policy three inconsistent ways — a structure never seen in training. Schema-constrained decoding would remove this failure class.
Judged by a lite-tier model. Stronger judges were quota-limited. Position inconsistency was 15–18%.
English only, small-business domain.

Licence

Apache-2.0, inheriting the base model's licence. The training dataset is MIT.