xevyo-01 / code /README.md
meet4150's picture
xevyo-01 round 4: merged weights, adapter, model card, code
a46f204 verified
|
Raw History Blame Contribute Delete
3.77 kB
# xevyo-01 code
The code that renders requests for xevyo-01, checks its plans, serves it, retrains it and scores
it. Training data and the evaluation sets are not included.
| Path | What it does | Needs |
|---|---|---|
| `spec/render.py` | Builds the exact system and user messages the model was trained on (`render_messages(example, window)`), and the canonical plan text (`render_target`) | Python standard library |
| `spec/validate.py` | Checks a plan: the JSON schema plus the rules a schema cannot express (offered tools, argument shapes, field conditions, directive preconditions) | `jsonschema` |
| `spec/*.json`, `spec/system_prompt.md`, `spec/examples.jsonl` | The prompt template, tool and directive catalogues, schemas and 20 worked examples the two files above read | |
| `spec/test_spec.py` | 69 tests for rendering and validation | `pytest` |
| `serve/vllm_schema.py` | The per-turn JSON schema for vLLM guided decoding (`response_format(offered_tools)`) | |
| `serve/hard_rules.py` | Safety and consistency rules to apply to every plan before acting on it (`apply(plan, example)`) | |
| `serve/merge_adapter.py` | Merges a LoRA adapter into the base weights (W + (alpha/r)·B·A) so vLLM can serve it | `torch`, `safetensors` |
| `train/sft_lora.py`, `train/configs/r4.yaml` | LoRA supervised fine-tuning, loss on the plan tokens only; the round-4 settings | `train/requirements.txt` |
| `train/Dockerfile` | The training image (PyTorch 2.10, CUDA 12.8, pinned stack) | Docker |
| `eval/run_eval.py`, `eval/score.py` | Sends rows to an endpoint, validates each answer and scores it field by field | `jsonschema` |
## Render a request and check a plan
```python
import json, sys
sys.path[:0] = ["code/spec", "code/serve"]
import render, validate, hard_rules
from vllm_schema import response_format
example = json.loads(open("code/spec/examples.jsonl", encoding="utf-8").readline())["input"]
messages = render.render_messages(example) # window="agent" for the longer history window
fmt = response_format(example["offered_tools"]) # pass as response_format to vLLM
# ... call the model with `messages` and `fmt`, then:
plan = json.loads(reply_text)
plan, changes = hard_rules.apply(plan, example)
errors = validate.validate_plan(plan, example) # [] when the plan is usable
```
`example` follows `spec/example_input.schema.json`: the message, the earlier turns, the tools on
offer and a few flags (image attached, memory on, locale).
## Context windows
- `trained` (the default): the last 8 messages; user turns up to 500 characters, answers up to 600.
- `agent`: newest first, up to 24 messages within 12,000 characters; user turns up to 2,000
characters, the latest answer up to 6,000, older answers up to 1,200.
Round 4 was trained on both. Use the window that matches how much history your application has.
## Retrain
```bash
docker build -t xv01-train code/train
# rows: {"input": <example_input>, "target": "<render_target(plan)>"} in data/train.jsonl and data/eval.jsonl
torchrun --nproc_per_node 2 code/train/sft_lora.py --config code/train/configs/r4.yaml
python code/serve/merge_adapter.py --adapter runs/xevyo01-r4/adapter_vllm --out models/xevyo-01
```
The image sets `HF_HUB_OFFLINE=1`; download `Qwen/Qwen3.5-9B` at the pinned revision into
`HF_HOME` first, or unset it.
## Score
```bash
python code/eval/run_eval.py --base http://localhost:8000 --model xevyo-01 \
--gold your_rows.jsonl --out results/run1
```
Rows: `{"id", "input": <example_input>, "target": "<canonical plan>"}`. The report gives
schema-valid %, exact plan %, safety failures and accuracy per field. `--traffic` (in `score.py`)
re-weights by a traffic mix you set in `TRAFFIC_WEIGHTS`.
## Licence
Apache 2.0, like the model.