README.md

metadata

license: apache-2.0
base_model: Qwen/Qwen3.5-4B
tags:
  - code
  - qlora
  - checkpoint-fabric
  - work-in-progress

AVA v3 — Training Checkpoints (work in progress)

Training artifacts for AVA v3.0, a coding-specialist model built on a $0 compute budget: free Colab/Kaggle GPU quota + one 4 GB-VRAM laptop, with Hugging Face Hub as the single source of truth for resume-anywhere training.

Recipe: QLoRA (r=16, all-linear) on Qwen/Qwen3.5-4B (native 3:1 Gated DeltaNet hybrid, 262K ctx), trained on nvidia/OpenCodeReasoning + bigcode/commitpackft (hash-anchored edit dialect), completion-only loss, decontaminated against the eval sets below.

Donor baseline (C1, the bar every checkpoint is gated against)

4-bit NF4, zero-shot, non-thinking, greedy — deployment-realistic protocol:

Benchmark	Score
HumanEval+ (164, executed)	67.68
MBPP+ (378, executed)	66.14
ARC-Easy (floor >= 75)	93.98
MMLU (floor >= 45)	55.50

Repo layout

• reports/c1_donor_baseline.json — immutable baseline (per-task results)
• reports/probes/ — mid-training probe evals (matched-subset deltas)
• checkpoints/C5/ — live training state: adapters + optimizer + RNG + data cursor; LATEST.json pointer written last (atomic resume)
• wheels/ — cached causal-conv1d builds per platform tag
• archive/ — forensic notes on reset runs

Status

• C1 donor baseline: done
• C5 SFT: run 2 in progress (run 1 reset after a data-cursor bug was caught by probe evals — see archive/)
• Gate: candidate must stay within 2pp of donor code scores and above the sanity floors, evaluated by the same harness that set the baseline

Training pipeline, evals and the resumable-notebook autopilot live in the AVA repo under experiments/exp6_v3/.