How to use from
Docker Model Runner
docker model run hf.co/43rt/claude-code-local-classifier:Q4_K_M
Quick Links

claude-code-local-classifier (Qwen3.5-4B LoRA)

ๅพฎ่ฐƒ Qwen3.5-4B ๅš Claude Code ็š„ๆœฌๅœฐๅฎ‰ๅ…จๅˆ†็ฑปๅ™จ๏ผˆprompt injection ๅ“จๅ…ต๏ผ‰ใ€‚ๆ•ฐๆฎไธๅ‡บๅŸŸใ€้˜ˆๅ€ผๆฆ‚็އๅฏ่ฐƒใ€ๆœฌๅœฐ GPU/CPU ๆŽจ็†ใ€‚298 ๆก holdout๏ผšdeny_recall 0.866 / false_block 0.005 / auc 0.945ใ€‚

ๅบ•ๅบงไธŽ่ฎญ็ปƒ

  • ๆ€่ทฏๆฅๆบ๏ผšๅŸบไบŽ Jev๏ผˆTypeSafe AI ็š„ๅฟซ้€ŸๆŽจ็†ๆจกๅž‹๏ผ‰โ€”โ€”ไธ็”Ÿๆˆๆ–‡ๆœฌใ€ๅชๅš็ป“ๆž„ๅŒ–ๅ†ณ็ญ–๏ผŒ่ฟ”ๅ›ž true/false ไบŒๅˆ†็ฑปๆฆ‚็އใ€‚ๆœฌๆจกๅž‹ๆฒฟๆญคๆ€่ทฏ๏ผš็บฏๅˆ†็ฑปใ€ไธ็”Ÿๆˆ็†็”ฑใ€‚
  • ๅบ•ๅบง๏ผšQwen/Qwen3.5-4B๏ผˆๅคšๆจกๆ€็‰ˆ๏ผŒ่ฎญ็ปƒๆ—ถ็”จ get_text_config() ๅ‰ฅ่ง†่ง‰็ผ–็ ๅ™จ๏ผŒๅชๅ–่ฏญ่จ€้ƒจๅˆ†๏ผ‰
  • LoRA๏ผšr=32๏ผŒ7 ็›ฎๆ ‡ๆจกๅ—๏ผˆq/k/v/o + mlp gate/up/down๏ผ‰๏ผŒadapter ~181M
  • ๆ–นๆณ•๏ผšRLCD ๅˆ†็ฑปโ€”โ€”ๆจกๅž‹ๅญฆ yes(deny)/no(allow) ๅ• token ๆฆ‚็އ๏ผŒไธ็”Ÿๆˆ reason
  • ่ฎญ็ปƒๆ•ฐๆฎ๏ผšๅˆๆˆ RLCD case + ็œŸๅฎž Claude Code ไผš่ฏ holdout๏ผˆ298 ๆก๏ผ‰ใ€‚ๅŽŸๅง‹ไผš่ฏๆ•ฐๆฎๅซ้š็ง๏ผŒไธๅœจๆœฌไป“ๅบ“ๅ‘ๅธƒ๏ผŒไป…ๆไพ› 6 ๆกๅˆๆˆๆ ทๆœฌๆผ”็คบๆ ผๅผ๏ผˆ่ง GitHub ไป“ๅบ“ datasets/samples.jsonl๏ผ‰
  • prompt ๆ ผๅผ๏ผšchatml ไธ‰ๆฎต๏ผˆpolicy_summary + <transcript>+action + ่ฃๅ†ณๆŒ‡ไปค๏ผ‰๏ผŒ่ฎญ็ปƒ/ๆŽจ็†/้ชŒๆ”ถๅŒไธ€็œŸๆบ prompt_assembly.py

ๆŽจ็†็”จๆณ•

ๆจกๅž‹ไธ่ฏป็”Ÿๆˆๆ–‡ๆœฌ๏ผˆQwen3.5 thinking ๅพช็Žฏไผš่พ“ๅ‡บ IMD๏ผ‰๏ผŒ่€Œๆ˜ฏ่ตฐ llama-server /completion ่ฏป yes/no logprob ๆฆ‚็އ๏ผš

# 1. ่ตท llama-server๏ผˆGGUF ็‰ˆ๏ผ‰
llama-server -m qwen35-clf-v2-q4_k_m.gguf --host 127.0.0.1 --port 11943 \
  -c 16384 --parallel 1 -ngl 999 -ctk q8_0 -ctv q8_0

# 2. /completion ่ฏปๆฆ‚็އ๏ผˆprompt ็”ฑ prompt_assembly.build_prompt ็”Ÿๆˆ๏ผ‰
curl 127.0.0.1:11943/completion -H 'Content-Type: application/json' -d '{
  "prompt": "<|im_start|>system\n{policy_summary}<|im_end|>\n<|im_start|>user\n<transcript>\n{\"user\":\"...\"}\n{\"Bash\":{...}}\n</transcript>\n\n{่ฃๅ†ณๆŒ‡ไปค}<|im_end|>\n<|im_start|>assistant\n",
  "n_predict": 1, "n_probs": 60, "temperature": 0, "cache_prompt": true
}'
# ๅ“ๅบ” completion_probabilities[0].top_logprobs ๅ– yes/no logprob
# p_yes = ฯƒ((lp_yes - lp_no) / T)๏ผŒT=1.0448๏ผˆoutput/adapter/temperature.json๏ผ‰
# block ๅˆคๅฎš๏ผšp_yes >= 0.5 => deny

ๅฎŒๆ•ด็คบไพ‹่„šๆœฌ่ง GitHub ไป“ๅบ“ eval/verify_local.pyใ€eval/eval_holdout_full.pyใ€‚

ๆŒ‡ๆ ‡๏ผˆ298 holdout๏ผ‰

ๆŒ‡ๆ ‡ ๆœฌๅœฐ v2 (RTX 4050 GPU) ๆœๅŠกๅ™จ็บฏ CPU transformers ่ฎญ็ปƒ็Žฏๅขƒ
deny_recall@0.5 0.866 0.866 0.884
false_block@0.5 0.005 0.005 0.005
auc 0.945 โ€” 0.979
sep 0.799 0.726 0.819
latency 1.8s/ๆก p50 33ms โ€”

ๆ–‡ไปถ

  • adapter/ โ€” LoRA adapter๏ผˆ~181M๏ผ‰๏ผŒๆŒ‚ๅˆฐ Qwen/Qwen3.5-4B ๅบ•ๅบง
  • qwen35-clf-v2-q4_k_m.gguf โ€” GGUF Q4_K_M๏ผˆ~2.6G๏ผ‰๏ผŒllama-server ็›ดๆŽฅ่ท‘
  • temperature.json โ€” ๆ กๅ‡†ๆธฉๅบฆ T=1.0448

ๅ…ณ้”ฎๅ‘๏ผˆๅคๅˆปๅฟ…่ฏป๏ผ‰

  • convert GGUF **ไธ่ฆๅธฆ --no-nextn**๏ผˆไผš็ณป็ปŸๆ€งๅŽ‹็ผฉ logits๏ผŒdeny p_yes 0.85โ†’0.47๏ผ‰ใ€‚ๆœฌ GGUF ๆ˜ฏ v2๏ผŒไธๅธฆใ€‚
  • ่ฎญ็ปƒ็”จ policy_summary.txt๏ผˆ15K ็ฒพ็ฎ€ๆ‘˜่ฆ๏ผ‰ไฝœ system๏ผŒไธ่ฆ็”จ่ฏทๆฑ‚ไฝ“็š„ๅฎŒๆ•ด rules๏ผˆ139K๏ผŒไธŽ่ฎญ็ปƒไธๅŒๆž„๏ผ‰ใ€‚
  • thinking ๅพช็Žฏ็ป•่ฟ‡ๆ–นๅผ = ่ฏป /completion ็š„ yes/no logprob๏ผŒไธ่ฏป็”Ÿๆˆๆ–‡ๆœฌใ€‚

่ฏฆ่ง GitHub ไป“ๅบ“ docs/pitfalls.mdใ€‚

ๅ…่ดฃ

  • ๆœฌๆจกๅž‹ๆ˜ฏ prompt injection ่พ…ๅŠฉๅ“จๅ…ต๏ผŒไธๆ›ฟไปฃๅฎŒๆ•ดๅฎ‰ๅ…จๅฎกๆŸฅใ€‚
  • ๅŽŸๅง‹่ฎญ็ปƒ/่ฏ„ๆต‹ๆ•ฐๆฎๅซ็œŸๅฎžไผš่ฏ๏ผŒไธๅœจๆœฌไป“ๅบ“ๅ‘ๅธƒ๏ผ›ไป…ๆไพ›ๅˆๆˆๆ ทๆœฌๆผ”็คบๆ ผๅผใ€‚
  • policy_summary ๆ˜ฏ้€š็”จ AI ๅฎ‰ๅ…จ่ง„ๅˆ™ๆ‘˜่ฆ๏ผ›ๅฎŒๆ•ด rules ไธ้šๆจกๅž‹ๅ‘ๅธƒใ€‚

็›ธๅ…ณ

  • GitHub ไป“ๅบ“๏ผˆ่ฎญ็ปƒ็ฎก็บฟ + ้ƒจ็ฝฒ + eval + router ๆ•ดๅˆ๏ผ‰๏ผš้“พๆŽฅๅพ…ๅกซ
  • ๆŠ€ๆœฏๅšๅฎข๏ผˆๅŠจๆœบ + ่ฟ‡็จ‹ + ่ธฉๅ‘ๅค็›˜๏ผ‰๏ผš้“พๆŽฅๅพ…ๅกซ
  • HuggingFace ๆœฌไป“ๅบ“๏ผšhttps://huggingface.co/nathaniel-min/claude-code-local-classifier
Downloads last month
55
GGUF
Model size
4B params
Architecture
qwen35
Hardware compatibility
Log In to add your hardware

4-bit

Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐Ÿ™‹ Ask for provider support

Model tree for 43rt/claude-code-local-classifier

Finetuned
Qwen/Qwen3.5-4B
Adapter
(1,222)
this model