heyimmitzu/glm-oss-8b-scripts

glm-oss-8b — training scripts

Scripts used to fine-tune Llama-3.1-8B-Instruct on the GLM-5.2-Conversation dataset with Unsloth + QLoRA (r=16).

Files

• train.py — QLoRA SFT via Unsloth + TRL. Trains, saves the LoRA adapter, merges to 16-bit, exports GGUF (Q8_0 / Q4_K_M), writes Ollama Modelfiles + README, and pushes everything to the Hub.
• watchdog.py — background monitor: logs hourly snapshots, watches disk space, and stops the RunPod pod via API when the job finishes.
• upload_to_hub.py — standalone helper to upload a model folder to the Hub.

Reproduce (example)

python3 train.py \
  --subset 20000 \
  --batch_size 4 --grad_accum 2 \
  --max_length 2048 --num_epochs 1 \
  --learning_rate 2e-4 --warmup_steps 100 \
  --grad_ckpt standard \
  --push_to_hub --repo_id heyimmitzu/glm-oss-8b \
  --gguf_quants q8_0,q4_k_m

Set your write token via HF_TOKEN or /workspace/.hf_token.