File size: 2,871 Bytes
c70c8ec 2128ec9 c70c8ec 71d5c9a c70c8ec 2128ec9 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 | ---
license: mit
base_model: inclusionAI/Ling-3.0-tiny
library_name: safetensors
---
# 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](https://huggingface.co/inclusionAI/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](https://huggingface.co/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. |