Modelpublic

cb1c7/psp1g-qwen3-4b-v15-hotfix

sourceHugging Faceotherupdated 4mo agoView on Hugging Face

0likes55downloads

Model Card

qwen3-4b-v15-hotfix

Selected keeper from the v3.1 Qwen3-4B hotfix

Recommended decoding:

json

{
  "temperature": 0.70,
  "top_p": 0.88,
  "top_k": 40,
  "repetition_penalty": 1.12,
  "max_new_tokens": 48
}

Avoid high-temperature serving for now. The model is intentionally short-form and style-preserving; higher heat makes old weird branches more likely.

Included artifacts:

—merged-hf/: merged FP16 Hugging Face model
—qwen3-4b-v15-hotfix-fp16.gguf
—qwen3-4b-v15-hotfix-q4_k_m.gguf: practical default
—qwen3-4b-v15-hotfix-q5_k_m.gguf: better quality/size tradeoff
—qwen3-4b-v15-hotfix-q8_0.gguf: high quality quant

Training metadata:

json

{
  "training_metrics.json": {
    "train_runtime": 70.5763,
    "train_samples_per_second": 10.202,
    "train_steps_per_second": 0.638,
    "total_flos": 1697659702149120.0,
    "train_loss": 1.347390800052219,
    "epoch": 0.16289592760180996,
    "wall_time_seconds": 71.3073878288269,
    "gpu": "NVIDIA GeForce RTX 3090",
    "bf16": true,
    "max_memory_allocated_gb": 10.476,
    "max_memory_reserved_gb": 11.906
  },
  "dataset_stats.json": {
    "train": {
      "rows": 4420,
      "token_length": {
        "min": 95,
        "p50": 102,
        "p90": 108,
        "p95": 108,
        "p99": 110,
        "max": 110
      },
      "assistant_tokens": {
        "p50": 12,
        "p90": 16,
        "max": 17
      },
      "assistant_words": {
        "p50": 7,
        "p90": 8,
        "max": 9
      }
    },
    "eval": {
      "rows": 209,
      "token_length": {
        "min": 92,
        "p50": 106,
        "p90": 121,
        "p95": 126,
        "p99": 131,
        "max": 182
      },
      "assistant_tokens": {
        "p50": 14,
        "p90": 17,
        "max": 53
      },
      "assistant_words": {
        "p50": 7,
        "p90": 11,
        "max": 18
      }
    }
  }
}