Status: superseded. Canonical coding model: scbe-coding-agent-vtc-qwen15-v1-gguf. Kept as research history.
scbe-bijective-tongue-coder-qwen-kaggle-v1
LoRA adapter on Qwen/Qwen2.5-Coder-0.5B-Instruct for the SCBE-AETHERMOORE bijective DSL / Sacred Tongues coding lane.
Status: pre-frozen-eval (do not promote)
This adapter is uploaded as a checkpoint, not a promoted lane winner. It has not yet cleared the SCBE executable-accuracy promotion gate (perplexity + executable accuracy on bijective_dsl_v1_holdout). Treat as a memorization checkpoint until that gate is cleared.
Two flags worth knowing about before reuse:
- Overfit risk. Training ran 260 steps over 37.16 epochs on a small bijective-DSL dataset. Training loss collapsed to 0.052 and training token-accuracy reached 98.2% — both warning signs of memorization on a dataset of this size.
- Pre-eval. No held-out evaluation has been recorded yet. A sibling round (
polly-auto-dsl-syn-v2) is still training; the better of the two will be the lane winner after frozen-eval.
Training metrics (from checkpoint-260/trainer_state.json)
| Metric | Value |
|---|---|
| Steps | 260 |
| Epochs | 37.16 |
| Loss start | 3.7675 |
| Loss end | 0.0522 |
| Loss drop | 3.7153 |
| Token accuracy start | 0.4318 |
| Token accuracy end | 0.9824 |
| Accuracy gain | 0.5506 |
| Max grad norm | 2.3050 |
| LR start | 7.9997e-05 |
| LR end | 3.1083e-09 (cosine fully decayed) |
Configuration
LoRA rank 16, alpha 32, dropout 0.05; targets up_proj down_proj o_proj q_proj gate_proj k_proj v_proj. PEFT 0.18.1.
Quick start
from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer
base = AutoModelForCausalLM.from_pretrained("Qwen/Qwen2.5-Coder-0.5B-Instruct")
model = PeftModel.from_pretrained(base, "issdandavis/scbe-bijective-tongue-coder-qwen-kaggle-v1")
tok = AutoTokenizer.from_pretrained("Qwen/Qwen2.5-Coder-0.5B-Instruct")
Provenance
- Platform: Kaggle (
issacizrealdavis/polly-auto-bijective-tongue-coder-v1) - Date: 2026-04-26
- Training framework: TRL SFTTrainer + PEFT LoRA
- Project: SCBE-AETHERMOORE (
https://github.com/issdandavis/SCBE-AETHERMOORE)
License
Apache-2.0 for the adapter weights. Base model (Qwen/Qwen2.5-Coder-0.5B-Instruct) is governed by its own Tongyi Qianwen license.
- Downloads last month
- 10
Model tree for issdandavis/scbe-bijective-tongue-coder-qwen-kaggle-v1
Base model
Qwen/Qwen2.5-0.5B