From 7351849a76f9961e7394e4dc2fe72b84030f4069 Mon Sep 17 00:00:00 2001
From: Kral <kral@local>
Date: Sat, 3 Oct 2026 18:23:39 +0200
Subject: [PATCH] Stage 1 step 0: train/.venv (py3.11, mlx-lm 0.32), MLX 4-bit
 Qwen3.8-27B, serve script, baseline runner, README

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_014aUaQeLnwbb1zTpN7kHeat
---
 train/README.md   | 54 ++++++++++++++++++++++++++++++++++++++++++
 train/baseline.py | 60 +++++++++++++++++++++++++++++++++++++++++++++++
 train/serve.sh    | 11 +++++++++
 3 files changed, 125 insertions(+)
 create mode 100644 train/README.md
 create mode 100644 train/baseline.py
 create mode 100755 train/serve.sh

diff --git a/train/README.md b/train/README.md
new file mode 100644
index 0000000..79f2c63
--- /dev/null
+++ b/train/README.md
@@ -0,0 +1,54 @@
+# Stage 1 training (train/)
+
+Task: `docs/stage1-training-task.md`. State of the work: this file and `train/STATE.md` (later steps).
+
+## Environment
+
+- `train/.venv`: Python 3.11.17 (created with `uv`, Homebrew; the system Python 3.9 is not changed).
+- `mlx-lm` 0.32.0, `mlx` 0.32.3.
+
+## Base model
+
+- Base model: `Qwen/Qwen3.8-27B` (architecture `qwen3_5`, 27.8B, dense; Apache 2.0). It is the same base
+  model as the Ollama model `qwen3.8-27b-32k` that the harness used before.
+- Weights: **`mlx-community/Qwen3.8-27B-4bit`** (Hugging Face), local path `~/models/Qwen3.8-27B-4bit`.
+  Quantization: MLX affine, 4 bit, group size 64. Size 16.1 GB. Downloaded 2026-10-03.
+- Why not the Ollama weights (`qwen3.8:27b-mlx`): they are NVFP4 (modelopt) with one global scale per
+  layer. `mx.quantized_matmul` has no global scale, so `mlx_lm` cannot load them without a re-quantization
+  (a different model). Kral approved the Hugging Face download (2026-10-03).
+
+## Serving (`train/serve.sh`)
+
+`mlx_lm.server` at `http://127.0.0.1:8080/v1` (OpenAI-compatible). Base model without adapter; after
+training the same script with `--adapter-path`.
+
+Settings, the same as the earlier Ollama runs (runs 103, 203):
+
+| Setting | Ollama run | mlx_lm.server |
+|---|---|---|
+| Thinking | on, Ollama default level `medium` | `--chat-template-args '{"enable_thinking": true, "reasoning_effort": "medium"}'` (template default would be `xhigh`) |
+| temperature | 0.2 (sent by the harness llm agent; overrides the Modelfile value 1) | 0.2 (sent by the agent; server default `--temp 0.2`) |
+| top_p / top_k / min_p | 0.95 / 20 / 0 (Modelfile) | `--top-p 0.95 --top-k 20 --min-p 0` |
+| presence / repeat penalty | 0 / 1 (neutral) | not set (neutral) |
+| Output limit | none (context `num_ctx` 32768) | `--max-tokens 32768` (server default would be 512) |
+| Context | 32768 | no fixed limit (memory) |
+
+Thinking in the earlier Ollama runs: verified from run 103 (a turn with 3480 completion tokens and about
+100 visible tokens).
+
+Tool-call test (2026-10-03): one request with the harness system prompt and the MCP tool schemas returned
+`sap_pull_source(objectType=INTF, objectName=ZIF_DEMO_CHECK)` in the OpenAI `tool_calls` format; the
+arguments parse as JSON. The reasoning comes in a separate `reasoning` field. The chat template uses the
+qwen3_coder XML tool format; `mlx_lm` parses it. First request: 86 s (prompt of 9.5k tokens).
+
+## Eval subset
+
+`train/subset.json` (copy: `runs/stage1/subset.json`): 25 tasks, balanced over the categories
+(A, B, C, E, F: 3 each; D, G, H, I, K: 2 each), with T01. Use the same list before and after training.
+
+## Baseline
+
+- Runner: `python3 train/baseline.py --label baseline` (one task at a time; results
+  `runs/stage1/baseline.json`, run directories `runs/stage1/baseline/`).
+- The earlier T01 score 41.7 (run 103) used the Ollama NVFP4 weights. The baseline of 2026-10-03 with the
+  MLX 4-bit weights is the new reference for stage 1.
diff --git a/train/baseline.py b/train/baseline.py
new file mode 100644
index 0000000..a628bf5
--- /dev/null
+++ b/train/baseline.py
@@ -0,0 +1,60 @@
+"""Stage 1 eval runs on the fixed subset (train/subset.json) with a local OpenAI-compatible server.
+
+  train/.venv not needed: python3 train/baseline.py --label baseline [--only T01] [--run-base 20000]
+
+One task at a time. Results: runs/stage1/<label>.json (written after each task).
+"""
+import argparse
+import json
+import os
+import sys
+
+ROOT = os.path.dirname(os.path.dirname(os.path.abspath(__file__)))
+sys.path.insert(0, ROOT)
+from harness.adt_client import load_env  # noqa: E402
+from harness.agents import LlmAgent  # noqa: E402
+from harness.runner import Runner  # noqa: E402
+
+MODEL = os.path.expanduser("~/models/Qwen3.8-27B-4bit")
+BASE_URL = "http://127.0.0.1:8080/v1"
+
+
+def main():
+    load_env(os.path.join(ROOT, ".env"))
+    ap = argparse.ArgumentParser()
+    ap.add_argument("--label", default="baseline")
+    ap.add_argument("--only", nargs="*")
+    ap.add_argument("--run-base", type=int, default=20000)
+    a = ap.parse_args()
+    subset = json.load(open(os.path.join(ROOT, "train", "subset.json")))["tasks"]
+    out_path = os.path.join(ROOT, "runs", "stage1", f"{a.label}.json")
+    os.makedirs(os.path.dirname(out_path), exist_ok=True)
+    res = json.load(open(out_path)) if os.path.exists(out_path) else {"model": MODEL, "tasks": {}}
+    runs_root = os.path.join(ROOT, "runs", "stage1", a.label)
+    for i, t in enumerate(subset):
+        tid = t["id"]
+        if (a.only and tid not in a.only) or tid in res["tasks"]:
+            continue
+        pool = os.path.join(ROOT, "tasks") if tid.startswith("T") else os.path.join(ROOT, "tasks_gen", "eval")
+        agent = LlmAgent(MODEL, BASE_URL)
+        try:
+            rep, run_dir = Runner(pool, runs_root).run(tid, agent, a.run_base + i)
+        except Exception as e:  # noqa: BLE001  one broken run must not stop the series
+            res["tasks"][tid] = {"error": str(e)[:300]}
+            json.dump(res, open(out_path, "w"), indent=1)
+            print(json.dumps({"task": tid, "error": str(e)[:300]}), flush=True)
+            continue
+        h = rep.get("hidden_tests") or {}
+        res["tasks"][tid] = {"category": t["category"], "object_type": t["object_type"],
+                             "score": (rep.get("score") or {}).get("total"), "parts": rep.get("score"),
+                             "gates": rep.get("gates"), "hidden": f"{h.get('passed')}/{h.get('total')}",
+                             "tool_calls": rep.get("tool_calls"), "seconds": rep.get("seconds"),
+                             "agent_seconds": rep.get("agent_seconds"), "final": (rep.get("final_report") or "")[:300],
+                             "run_dir": os.path.relpath(run_dir, ROOT)}
+        json.dump(res, open(out_path, "w"), indent=1)
+        print(json.dumps({"task": tid, "score": res["tasks"][tid]["score"], "hidden": res["tasks"][tid]["hidden"],
+                          "tool_calls": rep.get("tool_calls"), "seconds": rep.get("seconds")}), flush=True)
+
+
+if __name__ == "__main__":
+    main()
diff --git a/train/serve.sh b/train/serve.sh
new file mode 100755
index 0000000..f927369
--- /dev/null
+++ b/train/serve.sh
@@ -0,0 +1,11 @@
+#!/bin/sh
+# Serve the base model (no adapter) with mlx_lm.server. Settings: train/README.md.
+# Usage: train/serve.sh [--adapter-path PATH]
+cd "$(dirname "$0")/.."
+exec train/.venv/bin/mlx_lm.server \
+  --model "$HOME/models/Qwen3.8-27B-4bit" \
+  --host 127.0.0.1 --port 8080 \
+  --temp 0.2 --top-p 0.95 --top-k 20 --min-p 0 \
+  --max-tokens 32768 \
+  --chat-template-args '{"enable_thinking": true, "reasoning_effort": "medium"}' \
+  "$@"