ling3-tiny-coder-ckpts — work-in-progress (PAUSED)

Status: paused / unfinished. This repo stores intermediate artifacts from a personal experiment: fine-tuning Ling-3.0-tiny (7.9B Bailing-hybrid MoE, KDA+MLA) into a C++ competitive-programming coder on Kaggle TPU v5e-8 with LoRA adapters. Development is inactive — the items here are training artifacts, not a finished model. Layout may change if work resumes.

What is in this repo

• models/ contains the only directly usable artifacts: two early-SFT merged GGUF exports (Q4_K_M / Q8_0 / bf16 + chat template). These are from an early mid-SFT point and are noticeably weaker than later checkpoints (see evals).
• evals/ stores raw LiveCodeBench-v6 (60-problem sample) + OJBench result JSONs.
• latest.json / routing_counts.npz / llama-bin-cache.tar.gz are pipeline bookkeeping (resume pointer, MoE expert-usage counters, llama.cpp build cache).

Training lineage

Stage	Checkpoint	Steps	Data	Notes
SFT	sft-21750	21,750	R1 reasoning traces (codeforces-cots) + Magicoder thinking-off, ~178M tok, 2 epochs	completed
CPT	cpt-2903	2,903	raw C++ (Stack-v2) acclimation, 24M tok	completed; best eval to date
agentic (v22)	agentic-600	600	SWE-smith C++ patches	partial
calib	calib-19460	19,460	convergence-calibrated SFT (branch-filtered traces, loop removal), 4 epochs from cpt-2903 weights-only	completed 2026-09

The calib stage was an attempt to fix the dominant failure mode — on hard problems the model reasons correctly but never emits code (runs out of tokens inside reasoning loops). Data curation per arXiv:2604.01702 (branching-keyword filter, loop removal, step deletion) + 4× more epochs.

Eval snapshots (LLM-generated, exec-checked; LiveCodeBench v6 sample)

Checkpoint	pass@1	case pass rate	GGUF quant
sft-14000	43.3%	41.1%	Q4_K_M
sft-14000	26.7%	26.7%	Q8_0 (early broken run)
cpt-2903	51.7%	39.7%	Q4_K_M
calib-19460	48.3%	36.4%	Q4_K_M

Honest bottom line: easy problems ≈ solved (100% at cpt-2903), but the convergence-calibration pass did not lift overall pass@1 — hard-problem generation-length failures persisted. The experiment was paused here rather than extended (next planned step was a small DPO stage targeting looped generations).

Resuming / reusing

• Base model: inclusionAI/Ling-3.0-tiny
• Adapter checkpoints pair 1:1 with optstate_* files for exact resume.
• The two models/ GGUFs are the only turnkey artifacts; a merged GGUF of the later checkpoints was never produced.