joshelu/qwen3-4b-eventspec-martech-merged

Qwen3-4B EventSpec — MarTech Event Taxonomy Generator (merged)

Fine-tuned Qwen/Qwen3-4B-Instruct-2507 that converts free-form marketing tracking requests into clean, implementation-ready analytics event specifications as strict JSON.

This repository contains the fully merged weights (LoRA adapter merged into the base model), so it loads like any standard model — no PEFT/adapter step required.

• Base model: Qwen/Qwen3-4B-Instruct-2507
• LoRA adapter (pre-merge): joshelu/qwen3-4b-Instruct-eventspec-martech-sft
• Method: Supervised Fine-Tuning (SFT) + LoRA, then merged
• Task: marketing tracking request → strict-JSON EventSpec

Intended use

Give the model a plain-English marketing/analytics tracking request; it returns a single JSON object specifying the events to implement (event names, parameters, triggers, consent and deduplication requirements, QA criteria, risk flags, etc.). Useful for MarTech / analytics engineering, GA4 / GTM instrumentation planning, and taxonomy standardization.

⚠️ Prompt format (required for correct output)

The model was trained with a specific chat format. You must reproduce it or output quality degrades sharply.

System prompt (use verbatim):

You are EventSpec, an expert MarTech analytics engineer. You convert free-form marketing tracking requests into clean, implementation-ready analytics event specifications.

Given a marketing tracking request, respond with a SINGLE valid JSON object and nothing else: no prose, no markdown, no code fences. The JSON must be strictly parseable.

The specification captures: a concise `request_summary`; the `business_goal`; the `tracking_scope` (platforms, page_or_screen, user_action, conversion_type); and a list of `recommended_events`. Each recommended event defines `event_name` (snake_case), `event_description`, `platform`, `event_type`, `required_parameters`, `optional_parameters`, `trigger_condition`, `consent_requirements`, and `deduplication_requirements`.

Use consistent snake_case event and parameter names, follow analytics best practices (GA4/GTM conventions where relevant), and respect privacy/consent requirements. Output only the JSON object.

User message format:

Convert this marketing tracking request into a clean analytics event specification.

Request:
<your tracking request here>

Usage (transformers)

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

repo = "joshelu/qwen3-4b-eventspec-martech-merged"
tok = AutoTokenizer.from_pretrained(repo)
model = AutoModelForCausalLM.from_pretrained(repo, dtype=torch.bfloat16, device_map="auto")

SYSTEM_PROMPT = (
    "You are EventSpec, an expert MarTech analytics engineer. You convert free-form "
    "marketing tracking requests into clean, implementation-ready analytics event "
    "specifications.\n\n"
    "Given a marketing tracking request, respond with a SINGLE valid JSON object and "
    "nothing else: no prose, no markdown, no code fences. The JSON must be strictly "
    "parseable.\n\n"
    "The specification captures: a concise `request_summary`; the `business_goal`; the "
    "`tracking_scope` (platforms, page_or_screen, user_action, conversion_type); and a "
    "list of `recommended_events`. Each recommended event defines `event_name` "
    "(snake_case), `event_description`, `platform`, `event_type`, `required_parameters`, "
    "`optional_parameters`, `trigger_condition`, `consent_requirements`, and "
    "`deduplication_requirements`.\n\n"
    "Use consistent snake_case event and parameter names, follow analytics best "
    "practices (GA4/GTM conventions where relevant), and respect privacy/consent "
    "requirements. Output only the JSON object."
)
USER_PREFIX = "Convert this marketing tracking request into a clean analytics event specification."

request = ("A pharmacy app wants to track refill reminders, refill started, refill submitted, "
           "refill ready, and pickup completed, but no medication names or prescription numbers "
           "should be sent.")

messages = [
    {"role": "system", "content": SYSTEM_PROMPT},
    {"role": "user", "content": f"{USER_PREFIX}\n\nRequest:\n{request}"},
]
prompt = tok.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tok(prompt, return_tensors="pt").to(model.device)

out = model.generate(**inputs, max_new_tokens=2048, do_sample=False,
                     pad_token_id=tok.eos_token_id)
print(tok.decode(out[0][inputs["input_ids"].shape[1]:], skip_special_tokens=True))

Usage (Inference Endpoints / chat_completion)

from huggingface_hub import InferenceClient
client = InferenceClient("https://YOUR-ENDPOINT.endpoints.huggingface.cloud", token="hf_...")
resp = client.chat_completion(
    messages=[
        {"role": "system", "content": SYSTEM_PROMPT},
        {"role": "user", "content": f"{USER_PREFIX}\n\nRequest:\n{request}"},
    ],
    max_tokens=2048,
    temperature=0,
)
print(resp.choices[0].message.content)

Recommended generation settings

• temperature = 0 (greedy) — best for stable, strictly parseable JSON.
• max_new_tokens >= 2048 — specs are long; a lower limit will truncate the JSON mid-string.
• Parse the output with json.loads; retry with a higher token limit if parsing fails.

Output schema

The model produces a single JSON object. Core keys:

• request_summary — one-line summary of the request.
• business_goal — the measurement objective.
• tracking_scope — { platforms, page_or_screen, user_action, conversion_type }.
• recommended_events — list of events, each with event_name (snake_case), event_description, platform, event_type, required_parameters, optional_parameters, trigger_condition, consent_requirements, deduplication_requirements.

Richer examples in the training data also include implementation_notes, qa_criteria, open_questions, and risk_flags, which the model produces as appropriate.

Training

• Dataset: joshelu/martech-event-taxonomy-mapper-training-data (private).
  • The 100-row train split, minus every id appearing in the validation or test split → 80 training rows (held-out eval stays honest).
  • Eval during training used the validation split (10 rows).
  • Assistant targets are compact, strict JSON (json.dumps(output, separators=(",", ":"))), validated as parseable before training.
• Method: SFT + LoRA, then merged.
  • LoRA: r=16, alpha=32, dropout=0.05, target modules q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj.
  • 3 epochs, effective batch size 4 (per-device 1 × grad-accum 4), lr 2e-4, warmup ratio 0.05, cosine schedule, max length 2048, bf16, gradient checkpointing.
  • Hardware: a10g-small (HF Jobs).
• Metrics (final): eval loss ≈ 0.320, eval mean token accuracy ≈ 0.924.

Limitations

• Trained on a small (80-row) dataset — coverage is limited to the taxonomy and styles seen in training; unusual domains may produce weaker specs.
• Output is a strong draft, not a substitute for review by an analytics engineer, especially for consent/privacy and PII handling.
• With very low max_new_tokens, long specs will be truncated and fail JSON parsing — keep the limit high and validate the parse.