qwen3-0.6b-explain

Qwen3-0.6B distilled to write grounded resume โ†” job-fit explanations as JSON: a fit summary, matched and missing skills, and 2โ€“4 evidence quotes copied verbatim from the job posting. Teacher: Qwen3-14B, filtered with the checks of resume-job-matcher. Code, data pipeline and evaluation: explain-distill.

Results (360 held-out resumeโ€“job pairs)

Base Qwen3-0.6B This model
Strict pass (valid JSON, verbatim 5โ€“25-word quotes, correct missing skills) 19.7% 50.3%
Quote precision 75.6% 93.8%
Valid JSON 95.8% 98.9%

Paired difference +30.6 pts (95% CI 24.2โ€“36.9). The Q4_K_M GGUF scores within 0.6 pts of f16.

Files

  • gguf/qwen3-0.6b-explain-Q4_K_M.gguf (0.40 GB, recommended), gguf/qwen3-0.6b-explain-Q8_0.gguf
  • merged/: fp16 Transformers weights; adapter/: the LoRA adapter (r16/ฮฑ32, all linear layers)

Run (CPU, llama.cpp)

hf download srish327/qwen3-0.6b-explain gguf/qwen3-0.6b-explain-Q4_K_M.gguf --local-dir .
llama-server -m gguf/qwen3-0.6b-explain-Q4_K_M.gguf --jinja -c 4096 --port 8080

Send the same chat messages resume-job-matcher builds (jobmatch.explain.build_messages) to /v1/chat/completions with "chat_template_kwargs": {"enable_thinking": false}: the model was trained on that rendering. Speed: ~1.7 s per explanation on a T4, ~25 s on 4 CPU cores.

Limits

Trained on synthetic resumes (written by the teacher from real postings) and English tech job postings. About half of the outputs still fail the strict checks, mostly quotes just outside 5โ€“25 words; validate quotes against the posting before showing them, as resume-job-matcher does.

Downloads last month
27
GGUF
Model size
0.6B params
Architecture
qwen3
Hardware compatibility
Log In to add your hardware

4-bit

8-bit

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

Model tree for srish327/qwen3-0.6b-explain

Finetuned
Qwen/Qwen3-0.6B
Quantized
(573)
this model