Resume adapter + coding/security-heavy mix

train_securecoder.py

@@ -74,14 +74,14 @@ MIX: list[Source] = [
     Source("lockon/xlam-function-calling-60k", 10000, "xlam", "dataset",
            note="xLAM: query/answers/tools API-call pairs"),
     # ---- coding ----------------------------------------------------------
-    Source("ise-uiuc/Magicoder-OSS-Instruct-75K", 10000, "magicoder",
+    Source("ise-uiuc/Magicoder-OSS-Instruct-75K", 20000, "magicoder",
            note="self-instruct code problems + solutions"),
     # ---- cybersecurity ---------------------------------------------------
-    Source("Trendyol/Trendyol-Cybersecurity-Instruction-Tuning-Dataset", 8000, "sua",
+    Source("Trendyol/Trendyol-Cybersecurity-Instruction-Tuning-Dataset", 12000, "sua",
            note="security instruction tuning"),
-    Source("AlicanKiraz0/Cybersecurity-Dataset-Fenrir-v2.1", 5000, "sua",
+    Source("AlicanKiraz0/Cybersecurity-Dataset-Fenrir-v2.1", 8000, "sua",
            note="broad security Q&A"),
-    Source("Humanlearning/CyberSecurity_OWASP-sft-dataset", 3000, "messages",
+    Source("Humanlearning/CyberSecurity_OWASP-sft-dataset", 5000, "messages",
            note="OWASP / secure-coding SFT"),
     Source("MrClipperz134/CTF-Instruct", 3000, "io",
            note="CTF instruction/output"),
@@ -591,6 +591,19 @@ def load_model_and_tokenizer(args):
         load_in_4bit=not args.no_4bit,
     )
 
+    if args.resume_adapter:
+        # Continue an existing QLoRA instead of drawing a fresh one. A fresh
+        # adapter would forget the tool-calling already trained into it.
+        from peft import PeftModel
+
+        log.info("resuming adapter %s (trainable)", args.resume_adapter)
+        model = PeftModel.from_pretrained(model, args.resume_adapter, is_trainable=True)
+        trainable = sum(p.numel() for p in model.parameters() if p.requires_grad)
+        total = sum(p.numel() for p in model.parameters())
+        log.info("trainable parameters: %s (%.2f%% of the model)", f"{trainable:,}",
+                 100 * trainable / max(total, 1))
+        return model, tokenizer
+
     targets: Any = args.target_modules
     if isinstance(targets, str) and targets != "all-linear":
         targets = [t.strip() for t in targets.split(",") if t.strip()]
@@ -778,6 +791,8 @@ def parse_args(argv=None):
 
     p.add_argument("--base-model", default="Qwen/Qwen3-Coder-30B-A3B-Instruct")
     p.add_argument("--output-repo", default=None, help="Hub repo for the LoRA adapter")
+    p.add_argument("--resume-adapter", default=None,
+                   help="existing LoRA repo to keep training (does not init a fresh adapter)")
     p.add_argument("--merge-repo", default=None, help="optional Hub repo for a 16-bit merge")
     p.add_argument("--output-dir", default="securecoder-out")
     p.add_argument("--private", action="store_true", help="create Hub repos as private")