Spaces:
Sleeping
Sleeping
Fix refiner prompt and use validated Qwen 1.5B model
Browse files100 tests passed; frontend build and hygiene gate passed. Five-demo greedy CPU check: kept improvements increased from 0/5 to 3/5. Scorer pin retained.
- Dockerfile +2 -2
- README.md +2 -2
- server/prompt.py +1 -1
Dockerfile
CHANGED
|
@@ -13,8 +13,8 @@ RUN cd web && npm run build
|
|
| 13 |
|
| 14 |
FROM python:3.12-slim
|
| 15 |
ARG CODEHYGIENE_REF
|
| 16 |
-
ARG REFINER_REPO=Qwen/Qwen2.5-Coder-
|
| 17 |
-
ARG REFINER_FILE=qwen2.5-coder-
|
| 18 |
|
| 19 |
RUN apt-get update \
|
| 20 |
&& apt-get install -y --no-install-recommends libgomp1 git \
|
|
|
|
| 13 |
|
| 14 |
FROM python:3.12-slim
|
| 15 |
ARG CODEHYGIENE_REF
|
| 16 |
+
ARG REFINER_REPO=Qwen/Qwen2.5-Coder-1.5B-Instruct-GGUF
|
| 17 |
+
ARG REFINER_FILE=qwen2.5-coder-1.5b-instruct-q4_k_m.gguf
|
| 18 |
|
| 19 |
RUN apt-get update \
|
| 20 |
&& apt-get install -y --no-install-recommends libgomp1 git \
|
README.md
CHANGED
|
@@ -21,8 +21,8 @@ The scorer runs in the browser. The refiner runs here on the CPU. A patch is kep
|
|
| 21 |
| Variable | Default | What it does |
|
| 22 |
|---|---|---|
|
| 23 |
| `REFINER_BACKEND` | `llama` | `llama` runs the GGUF model. `mock` replays recorded patches. |
|
| 24 |
-
| `REFINER_REPO` | `Qwen/Qwen2.5-Coder-
|
| 25 |
-
| `REFINER_FILE` | `qwen2.5-coder-
|
| 26 |
| `REFINER_MODEL_PATH` | set by the Dockerfile | Path of the GGUF inside the image. |
|
| 27 |
| `CODEHYGIENE_REF` | pinned commit | Build arg. Version of the scorer and web assets. |
|
| 28 |
| `MAX_FILE_CHARS` | `12000` | Longer files are not sent to the model. |
|
|
|
|
| 21 |
| Variable | Default | What it does |
|
| 22 |
|---|---|---|
|
| 23 |
| `REFINER_BACKEND` | `llama` | `llama` runs the GGUF model. `mock` replays recorded patches. |
|
| 24 |
+
| `REFINER_REPO` | `Qwen/Qwen2.5-Coder-1.5B-Instruct-GGUF` | Build arg. Hub repo of the GGUF. |
|
| 25 |
+
| `REFINER_FILE` | `qwen2.5-coder-1.5b-instruct-q4_k_m.gguf` | Build arg. GGUF file name. |
|
| 26 |
| `REFINER_MODEL_PATH` | set by the Dockerfile | Path of the GGUF inside the image. |
|
| 27 |
| `CODEHYGIENE_REF` | pinned commit | Build arg. Version of the scorer and web assets. |
|
| 28 |
| `MAX_FILE_CHARS` | `12000` | Longer files are not sent to the model. |
|
server/prompt.py
CHANGED
|
@@ -11,7 +11,7 @@ def render_chatml(messages: list[dict[str, str]]) -> str:
|
|
| 11 |
return (
|
| 12 |
f"<|im_start|>system\n{system['content']}<|im_end|>\n"
|
| 13 |
f"<|im_start|>user\n{user['content']}<|im_end|>\n"
|
| 14 |
-
"<|im_start|>assistant\n
|
| 15 |
)
|
| 16 |
|
| 17 |
|
|
|
|
| 11 |
return (
|
| 12 |
f"<|im_start|>system\n{system['content']}<|im_end|>\n"
|
| 13 |
f"<|im_start|>user\n{user['content']}<|im_end|>\n"
|
| 14 |
+
"<|im_start|>assistant\n"
|
| 15 |
)
|
| 16 |
|
| 17 |
|