Push adapter at every checkpoint (--push-checkpoints) so cancel/timeout cannot lose a run

JSONL fallback for datasets without Json feature type

Add malware analysis + counter/remediation sources

Resume adapter + coding/security-heavy mix

Eval v8: restore out.append in code-sanity loop

Eval v7: code-sanity try-block covers chat template

Eval v6: restore _fetch_first_rows

Eval v5: BitsAndBytesConfig for MoE 4-bit

Eval v4: --merged-repo, eos_token_id, deps without unsloth

Eval v4: Hermes JSON-string tools, merged-repo loader, eos_token_id

Eval v3: Hermes conv support + longest-block AST

Eval v2: clean rewrite with PEP 723 deps + inlined helpers

Eval: inline helpers (no cross-script import)

Quantise: apt-install cmake before building llama-quantize

Fix PEP 723 blank line

Fix PEP 723 blank line

Quantise safetensors -> GGUF (llama.cpp path)

Merge only (GGUF in separate job)

Add eval script (tool-call validity, code sanity, security MCQ)

Add merge+GGUF script

Option B: --mix-scale, trackio-before-config fix

SecureCoder runbook

Robust Hub push + throughput reporting

Fix MoE router LoRA target + attention-only fallback

Drop dpevzner (config churn); mix ready

Format-adaptive tool rendering + resilient dataset loading

Fix xLAM tool schemas + dpevzner config

SecureCoder trainer v1

initial commit