Instructions to use cmndcntrlcyber/qwen14b-code-trainer-v9_mixed with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use cmndcntrlcyber/qwen14b-code-trainer-v9_mixed with PEFT:

from peft import PeftModel
from transformers import AutoModelForCausalLM

base_model = AutoModelForCausalLM.from_pretrained("Qwen/Qwen2.5-Coder-14B-Instruct")
model = PeftModel.from_pretrained(base_model, "cmndcntrlcyber/qwen14b-code-trainer-v9_mixed")

Transformers

How to use cmndcntrlcyber/qwen14b-code-trainer-v9_mixed with Transformers:

# Use a pipeline as a high-level helper
from transformers import pipeline

pipe = pipeline("text-generation", model="cmndcntrlcyber/qwen14b-code-trainer-v9_mixed")
messages = [
    {"role": "user", "content": "Who are you?"},
]
pipe(messages)

# Load model directly
from transformers import AutoModel
model = AutoModel.from_pretrained("cmndcntrlcyber/qwen14b-code-trainer-v9_mixed", device_map="auto")

Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use cmndcntrlcyber/qwen14b-code-trainer-v9_mixed with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "cmndcntrlcyber/qwen14b-code-trainer-v9_mixed"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "cmndcntrlcyber/qwen14b-code-trainer-v9_mixed",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/cmndcntrlcyber/qwen14b-code-trainer-v9_mixed

SGLang

How to use cmndcntrlcyber/qwen14b-code-trainer-v9_mixed with SGLang:

Install from pip and serve model

# Install SGLang from pip:
pip install sglang
# Start the SGLang server:
python3 -m sglang.launch_server \
    --model-path "cmndcntrlcyber/qwen14b-code-trainer-v9_mixed" \
    --host 0.0.0.0 \
    --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "cmndcntrlcyber/qwen14b-code-trainer-v9_mixed",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker images

docker run --gpus all \
    --shm-size 32g \
    -p 30000:30000 \
    -v ~/.cache/huggingface:/root/.cache/huggingface \
    --env "HF_TOKEN=<secret>" \
    --ipc=host \
    lmsysorg/sglang:latest \
    python3 -m sglang.launch_server \
        --model-path "cmndcntrlcyber/qwen14b-code-trainer-v9_mixed" \
        --host 0.0.0.0 \
        --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "cmndcntrlcyber/qwen14b-code-trainer-v9_mixed",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Docker Model Runner
How to use cmndcntrlcyber/qwen14b-code-trainer-v9_mixed with Docker Model Runner:
```
docker model run hf.co/cmndcntrlcyber/qwen14b-code-trainer-v9_mixed
```

Code-Trainer V9 — Qwen2.5-Coder-14B LoRA Adapter

A LoRA fine-tune of Qwen/Qwen2.5-Coder-14B-Instruct trained on the code-trainer-v9-mixed dataset. Designed for agentic code generation with structured tool calling via Hermes-format <tool_call> XML tags.

Part of the RTPI (Real-Time Pipeline Intelligence) project.

What's New in V9

V9 is a corrective release targeting <tool_call> tag emission reliability. The V8 model emitted correct tool names and JSON arguments but frequently omitted the <tool_call>/</tool_call> wrapper tags that Ollama and vLLM need for structured tool-call parsing. V9 applies five fixes:

Fix	Description	Impact
Tag density	Tool-calling examples increased from 31% to 42% of dataset (19K glaive + synthetic multi-call)	Stronger tag pattern learning
Stop-after-tag	All assistant turns with `</tool_call>` stripped of trailing text	Clean generation stop signal
SFTConfig	`dataset_text_field="text"`, `packing=False` prevents cross-example tag bleed	Sharper token boundaries
Curriculum	Two-phase training: 80% full dataset (lr=1e-4), then 20% tool-call-only polish (lr=2e-4)	Format reinforcement
Q5_K_M	Deployment quantization upgraded from Q4_K_M to Q5_K_M	Better weight fidelity for tag sequences

Evaluation Results

Tool Call Eval (V10 — 14 scenarios)

Pass rate: 92.9% (13/14) | Target: 80%

Scenario	Tool	Result
list_directory	LS	Pass
read_file	Read	Pass
run_command	Bash	Pass
search_code	Grep	Pass
find_files	Glob	Pass
write_file	Write	Pass
edit_file	Edit	Pass
fetch_url	WebFetch	Pass
track_todos	TodoWrite	Pass
invoke_skill	Skill	Pass
delegate_to_subagent	Task	Pass
check_scope	ScopeCheck	Pass
no_tool_needed	(none)	Fail — hallucinated a tool call
run_tests	Bash	Pass

Agent Eval (V10 — 7 multi-turn scenarios)

Progress rate: 100% (7/7) | Target: 5/7

All scenarios (read_and_analyze, find_and_fix_bug, run_and_debug, scope_aware_recon, skill_invocation_scan, delegate_recon, investigate_and_report) passed with tool usage, correct explanations, no refusals, and no loops.

Training Metrics

Metric	Value
Phase A eval_loss	0.4581
Phase B eval_loss (final)	0.4551
Phase A steps	2,020
Phase B steps	505
Phase B tool-call rows	25,957

Training Details

Dataset

cmndcntrlcyber/code-trainer-v9-mixed — 44,890 examples across 4 slices:

Slice	Source	Train	Description
Tool calling	glaiveai/glaive-function-calling-v2	17,125	Qwen-native `<tool_call>` format, multi-call synthetic examples
Agentic traces	greghavens/fable-5-coding-and-debugging-traces	8,994	Multi-turn coding/debugging with tool use
Code generation	cmndcntrlcyber/code-trainer-offsec-dataset	7,074	Offensive security code across 8 languages
Instruction	teknium/OpenHermes-2.5	7,208	English instruction-following (language anchor)

64.2% of training examples contain tool calls.

Curriculum Training

Two-phase SFT with trl SFTTrainer:

Phase A (80% of steps): Full mixed dataset at lr=1e-4 with cosine scheduler
Phase B (20% of steps): Tool-calling-only subset (25,957 rows) at lr=2e-4 for format polish

Hyperparameters

Parameter	Value
Base model	Qwen/Qwen2.5-Coder-14B-Instruct
LoRA rank (r)	32
LoRA alpha	64
LoRA dropout	0.05
Target modules	q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj
Effective batch size	16 (bs=1, grad_accum=16)
Max sequence length	4,096
Precision	BF16
Gradient checkpointing	Yes (use_reentrant=False)
Packing	False
Optimizer	AdamW (weight_decay=0.01)
Warmup ratio	0.05 (Phase A), 0.10 (Phase B)

Compute

Hardware: NVIDIA A100-80GB (HuggingFace Jobs)
Training time: ~20 hours (Phase A + Phase B)
Framework: transformers 5.7.0, trl 1.3.0, peft 0.19.1, torch 2.11.0+cu128

Usage

With PEFT + Transformers

from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

base = AutoModelForCausalLM.from_pretrained(
    "Qwen/Qwen2.5-Coder-14B-Instruct",
    dtype=torch.bfloat16,
    device_map="auto",
)
model = PeftModel.from_pretrained(base, "cmndcntrlcyber/qwen14b-code-trainer-v9_mixed")
tokenizer = AutoTokenizer.from_pretrained("cmndcntrlcyber/qwen14b-code-trainer-v9_mixed")

messages = [
    {"role": "system", "content": "You are a coding assistant with access to tools."},
    {"role": "user", "content": "List the files in the current directory."},
]
text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tokenizer(text, return_tensors="pt").to(model.device)
output = model.generate(**inputs, max_new_tokens=256)
print(tokenizer.decode(output[0], skip_special_tokens=False))

As GGUF (recommended for deployment)

See cmndcntrlcyber/qwen14b-code-trainer-gguf for merged + quantized Q5_K_M and Q4_K_M GGUF files ready for Ollama, llama.cpp, or vLLM.

Known Limitations

no_tool_needed scenario: The model occasionally hallucinated a tool call when no tool was needed (1/14 failure). The strong tool-call training bias makes the model eager to use tools even for pure knowledge questions.
Multilingual artifacts: Some responses still contain Unicode replacement characters at sequence boundaries (visible in eval previews as �). These are residual Qwen2.5 tokenizer artifacts that appear during greedy/sampling decoding.
Tag format inconsistency: While most responses use proper <tool_call> tags, some emit <toolcall> (no underscore). The nexus-harness fallback parser handles both variants.

Next Steps for Optimization

V10 tag normalization: Add a post-processing tokenizer constraint or logit bias that forces the model to emit <tool_call> (with underscore) consistently, eliminating the <toolcall> variant.
Negative examples for no-tool scenarios: Add 1-2K training examples where the user asks a knowledge question and the assistant responds with plain text (no tool call). This would fix the 1/14 hallucination failure.
Special token registration: Register <tool_call> and </tool_call> as special tokens in the tokenizer (single token IDs) rather than multi-token text sequences. This would make the tags quantization-proof.
Full 3-epoch training: V9 used 1 epoch with curriculum. A 3-epoch run with the same curriculum (Phase A 80% / Phase B 20% per epoch) may further reduce eval_loss.
DPO/ORPO alignment: Use the tool-call eval failures as negative examples in a preference optimization pass to reduce hallucinated tool calls.
Context length scaling: Current max_seq_length is 4096. Testing at 8192 with RoPE scaling could improve multi-turn agent performance.

Project

Repository: cmndcntrlcyber/code-trainer-pipeline
GGUF models: cmndcntrlcyber/qwen14b-code-trainer-gguf
Dataset: cmndcntrlcyber/code-trainer-v9-mixed

Framework Versions

PEFT 0.19.1
Transformers 5.7.0
TRL 1.3.0
PyTorch 2.11.0+cu128

Downloads last month: 45

Model tree for cmndcntrlcyber/qwen14b-code-trainer-v9_mixed

Base model

Qwen/Qwen2.5-14B

Finetuned

Qwen/Qwen2.5-Coder-14B