ZDDC/training-data

ZDDC Training Data Pipeline

Adaptive LoRA fine-tuning for Qwen3-Coder-Next (vllm/RedHatAI/Qwen3-Coder-Next-NVFP4).

Training data is collected reactively: when Qwen struggles on a task and a stronger model (Sonnet/Opus) is consulted, that interaction is captured as a training example. Over time, domain-specific LoRA adapters are trained to patch Qwen's weak spots.

---

Directory Layout

training-data/
├── raw/
│   └── interactions.jsonl      # All captured weak-spot interactions
├── processed/
│   ├── all.jsonl               # Deduplicated, combined dataset
│   ├── multi-domain.jsonl      # All domains merged
│   └── <domain>.jsonl          # Per-domain splits (auto-generated)
├── validation/
│   ├── train.jsonl             # 80% split
│   ├── val.jsonl               # 10% split
│   └── test.jsonl              # 10% split (never used during training)
├── adapters/
│   └── <domain>-lora-v1/       # Trained LoRA adapter
│   └── <domain>-lora-v1-merged/ # Merged standalone model
├── snapshots/
│   └── v<date>/                # Versioned dataset snapshots
├── collect-interaction.js      # Capture a weak-spot interaction
├── process.sh                  # Cluster raw data into domain splits
├── validate.sh                 # Check data quality before training
├── train.sh                    # Train a LoRA adapter
└── deploy.sh                   # Merge adapter into standalone model

---

Workflow

Step 1 — Collect a weak-spot interaction

When Qwen gets stuck or you ask it to consult Sonnet/Opus:

node collect-interaction.js \
  --query  "How do I parse ZDDC filenames?" \
  --qwen   "[Qwen's suboptimal answer]" \
  --expert "[Sonnet's correct answer]"

Optionally specify domain explicitly (otherwise auto-detected):

node collect-interaction.js \
  --query  "..." \
  --qwen   "..." \
  --expert "..." \
  --domain zddc-naming

Raw interaction is appended to raw/interactions.jsonl.

Step 2 — Process (after ~50 new interactions)

bash process.sh

Deduplicates, clusters by domain, creates train/val/test splits.

Step 3 — Validate

bash validate.sh

Checks JSONL validity, domain balance, and split sizes.

Step 4 — Train

bash train.sh              # train multi-domain adapter
bash train.sh zddc-naming  # train domain-specific adapter

Outputs LoRA adapter to adapters/<domain>-lora-v1/.

Step 5 — Deploy (optional)

bash deploy.sh             # merge multi-domain adapter
bash deploy.sh zddc-naming # merge specific adapter

Merges the LoRA weights into the base model and saves a standalone model.

---

Training Data Format

Each line in a .jsonl file is one training example:

{
  "messages": [
    {"role": "user",      "content": "Query that exposed weakness"},
    {"role": "assistant", "content": "Qwen's original response"},
    {"role": "user",      "content": "consult Sonnet"},
    {"role": "assistant", "content": "Expert's correct response"}
  ],
  "metadata": {
    "domain":         "zddc-naming",
    "adapter":        "lora-v1-zddc_naming",
    "timestamp":      "2025-10-31T14:30:00.000Z",
    "interaction_id": "int_1735648200000_abc123",
    "source":         "manual-expert-consultation"
  }
}

---

Auto-Detected Domains

Domain	Trigger keywords
zddc-naming	zddc, trackingnumber, revision, status code
html-architecture	html, spa, single-file, es module, vanilla js
build-system	build.sh, dist/, template.html
coding-debugging	debug, error, fix, console
reasoning-architecture	reason, analyze, architecture, design
general-coding	(default)

---

LoRA Configuration

Parameter	Value	Notes
Base model	Qwen/Qwen2.5-7B-Instruct	Replace with Qwen3 when available on HF
Rank	64	Increase to 128 if underfitting
Alpha	64	1:1 with rank
Target modules	q_proj, v_proj, k_proj, o_proj	All attention projections
Dropout	0.05	Light regularisation
Learning rate	1e-4	Cosine decay with 10% warmup
Epochs	3	Monitor val loss to catch overfitting
Batch size	8 effective	4 per-device × 2 gradient accumulation
Precision	bfloat16	Requires Ampere GPU or newer

---

Hardware Requirements

Setup	Min VRAM	Method	Notes
Qwen-7B LoRA	24 GB	LoRA bf16	Recommended
Qwen-7B QLoRA	16 GB	QLoRA 4-bit	Add --load_in_4bit flag
Qwen-14B LoRA	48 GB	LoRA bf16	Better quality

Your system has 96 GB VRAM — full LoRA on Qwen-14B is feasible.

---

When to Retrain

• Every 50–100 new interactions collected
• When a new domain accumulates 200+ examples
• After a major project phase where Qwen struggled repeatedly