qwen14b-code-trainer-v6-aggressive-full3

LoRA adapter for Qwen/Qwen2.5-Coder-14B-Instruct, trained for 3 epochs over an 8 K-row slice of the code-trainer-offsec-dataset as the Phase 4B follow-up to the Phase 4A aggressive sweep winner.

Status: kept on the Hub for transparency, but not the canonical Phase 5 conversion target. The 1-epoch / full-data sibling qwen14b-code-trainer-v6-aggressive outperformed this adapter on the full validation split (eval_loss 0.4724 vs 0.5126) — see the comparison below.

Intended use

Same as the Phase 4A sibling — instruction-following code generation across the 8 dataset languages. Use this adapter only if you specifically want to study the 3-epoch / sliced-data behaviour.

Training data

• Dataset: cmndcntrlcyber/code-trainer-offsec-dataset (revision main, text-only chat format).
• Splits seen: 8,000 train rows (--train-limit 8000 slice of 26,126), 500 val rows (--val-limit 500 slice of 3,265).
• Total samples seen: 24,000 (3 epochs × 8 K). The Phase 4A run saw 26,126 samples in 1 epoch.

Training procedure

Knob	Value
Base model	Qwen/Qwen2.5-Coder-14B-Instruct
Adapter	LoRA (PEFT), r = 64, alpha = 128, dropout = 0.05
Learning rate	3e-4 (cosine decay, warmup ratio 0.03)
Batch size × grad accum	4 × 4 (effective batch = 16)
Epochs	3
Sequence length	2,048
Precision	bfloat16 + gradient checkpointing
Hardware	HF Skills a100-large
Frameworks	transformers, peft, trl (SFTTrainer)
Job runtime	4 h 53 m (COMPLETED)
HF Job	69f8188e9d85bec4d76f1c5e

Evaluation — Phase 4A vs Phase 4B (full val 3,265 rows)

Metric	Phase 4A aggressive	Phase 4B aggressive-full3 (this)
eval_loss (full val)	0.4724	0.5126
eval_loss (500-row slice during training)	—	0.5102
Total samples seen	26,126	24,000
Epochs	1	3

The 500-row training-time eval already pointed in this direction; the full-val eval confirmed it. Conclusion: more unique examples beats more passes for this dataset and this base model. The 3-epoch slice produced a measurably weaker adapter despite seeing roughly the same compute budget.

See docs/sweep/phase4b-summary.md for the apples-to-apples writeup.

Limitations

• Worse than qwen14b-code-trainer-v6-aggressive on the same val split — use that adapter unless you need to reproduce this experiment.
• All other limitations from the Phase 4A card carry over (no safety tuning, multilingual eval approximate, adapter-only).

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-v6-aggressive-full3"

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()

Reproducibility

• Code: github.com/cmndcntrlcyber/code-trainer-offsec-pipeline
• Launch command:

  python -m src.phase4_qwen_finetuning.scripts.launch_full_training \
      --config src/config/v6_config.yaml --best-config aggressive \
      --train-limit 8000 --val-limit 500 --wait
  

• Cost: ~$15.60 on a100-large (4 h 53 m).