qwen14b-code-trainer-v7_mixed

qwen14b-code-trainer-v7_mixed

LoRA adapter for Qwen/Qwen2.5-Coder-14B-Instruct, fine-tuned on the code-trainer-v7-mixed dataset. This is the V7 corrective-action release — the first mixed-capability training run after V6's distributional collapse wiped tool-calling and agent behaviour from the model.

Part of the Code-Trainer / RTPI pipeline (GitHub).

Status: superseded by qwen14b-code-trainer-v8_mixed, which fixes multilingual hallucination and tool-call format mismatches discovered during V7 validation. Kept on the Hub for reproducibility.

Motivation — V6 distributional collapse

The V6 adapter was trained on 26K single-turn code-transcription examples (100% "extract the code from this screenshot"). This created three capability gaps:

1. No tool calling — zero training examples with function/tool calls
2. No multi-turn reasoning — every example was one-shot input→output
3. No complex instruction following — every system prompt was identical

V7 corrects this by blending three capability slices into a single mixed dataset.

Training data

• Dataset: cmndcntrlcyber/code-trainer-v7-mixed
• Total: 28,862 train / 3,206 validation (90/10 split, seed 42)

• Tool coverage: 18,413 / 28,862 train rows (63.8%) contain tool definitions
• Format: Unified ChatML with Hermes-style <tool_call> XML tags, compatible with Qwen2.5's native chat template.

Training procedure

Key changes from V6 (aggressive):

• lora_r: 32 (was 64) — less aggressive LoRA to reduce capability overwrite
• learning_rate: 1.5e-4 (was 3e-4) — gentler to preserve tool-calling circuits
• max_seq_length: 8192 (was 2048) — agent prompts + tool specs need more context

Validation

V7 validation tested all three capabilities via separate HF Jobs:

Known issues (fixed in V8)

1. Multilingual hallucination — the model appends Thai, Russian, or Chinese garbage tokens at sequence boundaries after tool-call JSON. Root cause: Qwen2.5's large multilingual vocabulary + low confidence at end-of-generation.
2. Tool-call format mismatch — Slice B (Hermes format) uses <tool_call> tags tokenized as regular text tokens, not the native Qwen2.5 tool-call template produced by apply_chat_template(tools=...). The model learned the JSON payload but not the exact token sequence Ollama expects.

How to use

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

base_id = "Qwen/Qwen2.5-Coder-14B-Instruct"
adapter_id = "cmndcntrlcyber/qwen14b-code-trainer-v7_mixed"

tokenizer = AutoTokenizer.from_pretrained(base_id)
model = AutoModelForCausalLM.from_pretrained(
    base_id, dtype=torch.bfloat16, device_map="auto",
)
model = PeftModel.from_pretrained(model, adapter_id)
model.eval()

messages = [
    {"role": "system", "content": "You are a coding assistant with tool access."},
    {"role": "user", "content": "Read the file main.py and summarise its structure."},
]
inputs = tokenizer.apply_chat_template(
    messages, return_tensors="pt", add_generation_prompt=True,
).to(model.device)
out = model.generate(inputs, max_new_tokens=512, do_sample=False)
print(tokenizer.decode(out[0][inputs.shape[1]:], skip_special_tokens=True))

Reproducibility

• Code: github.com/cmndcntrlcyber/code-trainer-pipeline
• Dataset build:

  python -m src.phase2_preprocessing.scripts.build_v7_mixed_dataset \
      --config src/config/config.yaml
  

• Training launch:

  python -m src.phase4_qwen_finetuning.scripts.launch_validation_sweep \
      --config src/config/config.yaml --only v7_mixed --wait
  

• Validation launch:

  python -m src.phase4_qwen_finetuning.scripts.launch_v7_validation \
      --config src/config/config.yaml --wait
  

• W&B project: rtpi-phase4-qwen14b