skip to content
← featured work
Research release complete Completed research · Fine-tuning v0.1 research package

CGAF-Tune

Conflict-Gated Adaptive Fine-Tuning for capability-preserving LLM adaptation.

GitHub repository ↗Hugging Face model ↗
OVERVIEW

What it is

CGAF measures domain and retention gradients by adapter group, then smoothly projects only the harmful component when gradients conflict. It evaluates this hypothesis against matched LoRA, rehearsal and hard-projection baselines.

ℹSTATUS BOUNDARY
The implementation and reproducibility release are complete; the conclusion remains deliberately cautious.
PROBLEM

Why it exists

Parameter-efficient fine-tuning can improve a narrow domain while degrading general capability. The research asks whether intervening only on conflicting gradients improves the adaptation–retention trade-off.

ARCHITECTURE

System flow

Domain batch adaptation Anchor batch retention Conflict measure per group Smooth gate local intervention LoRA update base frozen Matched eval trade-off
Figure 1 · Two-pass gradient engine with layer-local conflict measurement and gated projection.
CAPABILITIES

What is implemented

▸ Stable per-group smooth gradient projection
▸ PEFT-compatible LoRA parameter grouping
▸ Qwen3 integration with LoRA and 4-bit QLoRA
▸ Per-step diagnostics and adapter checkpoints
▸ Matched LoRA, rehearsal and hard-projection baselines
▸ Multi-seed forgetting and Pareto analysis
▸ Leakage audit and per-example evaluation
▸ Published adapter, model card and reproducibility package
TECHNOLOGY STACK
Python 3.10+PyTorchTransformersPEFTQwen3LoRAQLoRAHugging Facepytest
ENGINEERING

Key decisions

ADR-01

Falsifiable claim

accepted

Mixed or negative evidence is reported directly instead of being turned into a product claim.

ADR-02

Matched budgets

accepted

Baselines use controlled data, update and seed budgets.

ADR-03

Frozen base model

accepted

PEFT adapters are saved separately so the base model remains unchanged.

TRUST

Security, privacy and quality controls

  • ◆Validated JSONL and explicit evaluation schemas
  • ◆Seeded experiment matrix with checkpointed manifests
  • ◆Training/evaluation leakage audit
  • ◆Inspectable example-level predictions and intervals
  • ◆Model card separates evidence, limitations and intended use
EVIDENCE

Verified results

  • ✓Controlled-conflict, natural-data, ablation, long-adaptation and Phase 6 experiments completed
  • ✓CGAF achieved the highest mean domain F1 in the reported Phase 6 comparison
  • ✓CGAF did not improve balanced harmonic F1 over LoRA or rehearsal across three seeds
  • ✓Manuscript, statistical analysis, citation and reproducibility checklist published
  • ✓Qwen3-0.6B seed-42 adapter published on Hugging Face
CURRENT LIMITATIONS

What this does not claim

▲ACCURACY BOUNDARY
  • • The evidence does not support a state-of-the-art claim.
  • • Results are scoped to documented models, datasets, seeds and compute.
  • • Broader external replication is needed before generalisation.
NEXT

Roadmap

  • →Invite independent reproduction and critique
  • →Test broader models and data compositions if justified
  • →Keep future claims tied to versioned evidence