You can deploy this model and start earning money today!

issdandavis/
scbe-coding-agent-vtc-qwen15-v1-gguf

Name: scbe-coding-agent-vtc-qwen15-v1-gguf
Availability: InStock

Views

No views yet

Status: canonical. Current production candidate for the SCBE coding-agent lane. Held-out eval (n=6, identical harness): 5/6 functional, exact_output_only 0/6 -> 6/6 against base. Other scbe-coding-* repos in this account are research history and are tagged superseded.

SCBE Coding Agent — Qwen2.5-Coder-1.5B (VTC SFT) — GGUF

LoRA SFT of Qwen/Qwen2.5-Coder-1.5B-Instruct on 4,286 verified SCBE coding-agent trajectories (scbe-coding-agent-sft-full-coding-system-v8). Adapter: issdandavis/scbe-coding-agent-vtc-qwen15-v1.

Eval gate (held-out generic coding, n=6, identical harness across models)

model	functional	format	exact_output_only
base-1.5b	5/6	6/6	0/6
vtc-qwen05-v1	3/6	6/6	6/6
vtc-qwen15-v1	5/6	6/6	6/6

The fine-tune matches base correctness (5/6) while learning the bare-code output contract: exact_output_only 0/6 → 6/6 (base wraps every answer in markdown fences; the fine-tune emits pure code). This is the agent-relevant win — consistent with the Superficial Alignment Hypothesis (capability from pretraining, format/contract from SFT).

Files

scbe-coding-agent-vtc-qwen15-v1.Q4_K_M.gguf — Q4_K_M, recommended for Ollama on ≈12GB RAM
scbe-coding-agent-vtc-qwen15-v1.Q5_K_M.gguf — Q5_K_M, higher quality
scbe-coding-agent-vtc-qwen15-v1.Q8_0.gguf — near-lossless
Modelfile — points at Q4_K_M

Ollama

bash

hf download issdandavis/scbe-coding-agent-vtc-qwen15-v1-gguf --local-dir scbe-qwen15
cd scbe-qwen15 && ollama create scbe-coder -f Modelfile
ollama run scbe-coder "Write a Python function add(a, b). Return ONLY the code."