qwen14b-code-trainer-gguf

GGUF quantizations of the Code-Trainer fine-tuned model. The full adapter chain — DAPT (qwen14b-dapt-offsec), V9 SFT (qwen14b-code-trainer-v9_mixed), and V10 GRPO (qwen14b-code-trainer-v10-grpo) — is merged into Qwen/Qwen2.5-Coder-14B-Instruct and quantized via llama.cpp.

This is Phase 5 of the Code-Trainer / RTPI pipeline. The conversion runs as an HF Job on a100-large — the GPU sits idle, we use that flavor only for its 144 GB system RAM during the float16 merge step.

Files

Additional quantizations (Q8_0, F16) can be produced by passing --quants to launch_convert.py.

Intended use

• Local inference via llama-cli, llama-server, Ollama, LM Studio, or text-generation-webui.
• Phase 6 hot-swap target for the project's vLLM + Qwen-Agent stack — swapped in for compiled-language tasks alongside a smaller primary model.
• Out of scope: anything the upstream qwen14b-code-trainer-aggressive card flags as out of scope (no safety tuning, no non-code tasks).

Source

The GGUF is produced by merging the full adapter chain in order, then quantizing the merged model:

Qwen/Qwen2.5-Coder-14B-Instruct
  → merge DAPT LoRA    (qwen14b-dapt-offsec)
  → merge V9 SFT LoRA  (qwen14b-code-trainer-v9_mixed)
  → merge V10 GRPO LoRA (qwen14b-code-trainer-v10-grpo)
  → convert_hf_to_gguf.py + llama-quantize → Q5_K_M

Evaluation

Quality is inherited from the source adapter chain. The current source is V10 GRPO — an RL-tuned adapter trained on a rule-based tool-call formatting reward (mean reward ~0.14, see the V10 model card). The SFT foundation is V9 (40,401-row curriculum dataset, 64.3% tool coverage — see the V9 model card). Quantization to Q5_K_M typically introduces minimal perplexity penalty (< 1 %) for 14 B models; Q4_K_M introduces ~1–3 %.

Quick start

llama-server

llama-server \
  -m Qwen2.5-Coder-14B-Instruct-Q5_K_M.gguf \
  --host 0.0.0.0 --port 8080 \
  --ctx-size 8192 --n-gpu-layers 999

Ollama Modelfile

FROM ./Qwen2.5-Coder-14B-Instruct-Q5_K_M.gguf
TEMPLATE """{{ if .System }}<|im_start|>system
{{ .System }}<|im_end|>
{{ end }}{{ range .Messages }}{{ if eq .Role "user" }}<|im_start|>user
{{ .Content }}<|im_end|>
{{ else if eq .Role "assistant" }}<|im_start|>assistant
{{ .Content }}<|im_end|>
{{ else if eq .Role "tool" }}<|im_start|>tool
{{ .Content }}<|im_end|>
{{ end }}{{ end }}<|im_start|>assistant
"""
PARAMETER stop "<|im_start|>"
PARAMETER stop "<|im_end|>"
PARAMETER num_ctx 8192

llama-cpp-python

from llama_cpp import Llama

llm = Llama(
    model_path="Qwen2.5-Coder-14B-Instruct-Q5_K_M.gguf",
    n_ctx=8192,
    n_gpu_layers=999,
)
print(llm.create_chat_completion(messages=[
    {"role": "user", "content": "Write a Go function that reverses a UTF-8 string."},
])["choices"][0]["message"]["content"])

Limitations

• Lossy quantization. Q4_K_M is a 4-bit-mixed format; expect minor degradation vs. the unquantized adapter on long-form code. Q5_K_M is recommended for tool-calling workloads.
• No safety tuning. Inherits all caveats from the source adapter.
• Two quants shipped. Q5_K_M (recommended) and Q4_K_M (fallback). For Q8_0 / F16, regenerate with python -m src.phase5_deployment.scripts.launch_convert --quants Q8_0.

Reproducibility

set -a && source .env && set +a
python -m src.phase5_deployment.scripts.launch_convert \
    --config src/config/config.yaml --wait

• Code: github.com/cmndcntrlcyber/code-trainer-pipeline (src/phase5_deployment/)
• Config: src/config/pipeline-50.yml (deployment section)
• Cost: ~$2 on a100-large once the job runs.