Views
No views yet
| File | Size | Description |
|---|---|---|
chitta-hint-qwen-v5-model-b-q4_k_m.gguf | 851 MB | Active — v5 fine-tune, Qwen3-0.6B Q4_K_M |
chitta-hint-qwen-q4_k_m.gguf | 379 MB | Prior generation (v4) |
| Model | TP rate | TN rate | Quality | Overall |
|---|---|---|---|---|
| v5-model-b (this) | 76% | 100% | 0.76 | 88% |
| v4 (prior) | 35% | 100% | 0.67 | 68% |
hint_corpus_v5.jsonl — 3,048 (input, hint) pairs in ChatML formatfinetune_hint_v5_model_b.py — unsloth + TRL SFTTrainer, LoRA r=16gen_hint_corpus_live.py (from session transcripts), gen_hint_corpus_ssl.py (SSL distillation)chitta-hint-qwen-v5-model-b-q4_k_m.gguf (or symlink) at:~/.claude/models/chitta-hint-qwen-q4_k_m.ggufCHITTA_HINT_MODEL=PATH.1from llama_cpp import Llama
2
3llm = Llama(
4 model_path="chitta-hint-qwen-v5-model-b-q4_k_m.gguf",
5 n_ctx=512, n_gpu_layers=0
6)
7
8SYSTEM = ("Extract a single concise retrieval hint from the message. "
9 "Cover: personal preferences, tech choices, developer workflows, "
10 "domain expertise, project facts. "
11 "If nothing factual is present, output nothing.")
12
13prompt = (f"<|im_start|>system\n{SYSTEM}<|im_end|>\n"
14 f"<|im_start|>user\nI always use neovim, never vscode<|im_end|>\n"
15 f"<|im_start|>assistant\n")
16
17out = llm(prompt, max_tokens=80, temperature=0.0, stop=["<|im_end|>"])
18print(out["choices"][0]["text"].strip())
19# → User uses neovim exclusively, never VSCode.