Add files using upload-large-folder tool

MERGE_SEQUENCE.sh

@@ -0,0 +1,426 @@
+#!/usr/bin/env bash
+set -euo pipefail
+
+# ============================================================
+# RUNME.sh (self-contained, auto-discovers checkpoints)
+#
+# Modes:
+#   - Default (WATCH=0): snapshot checkpoints once and process them.
+#   - WATCH=1          : keep polling for new checkpoint-* dirs and process new ones.
+#
+# Output:
+#   /workspace/v126rc_exp3/F_r10000/checkpoint-*/residued
+# Logs:
+#   /workspace/v126rc_exp3/F_r10000/checkpoint-*/residued/merge.log
+#
+# RAM safety:
+#   Each checkpoint merge runs in a fresh Python process.
+#
+# Auto-stop (WATCH=1):
+#   If no new checkpoints appear for IDLE_LIMIT_SECONDS (default 600),
+#   the script exits and (optionally) deletes the RunPod pod.
+#
+# Robustness:
+#   If a checkpoint is incomplete and merge fails, we log it and retry later
+#   (WATCH=1 keeps running; WATCH=0 continues to next checkpoint).
+#
+# Cleanup:
+#   After a SUCCESSFUL merge, delete everything inside checkpoint-* except residued/
+# ============================================================
+
+# ---------------- CONFIG ----------------
+LR_DIR="${LR_DIR:-/workspace/Llama-3.2-3B-Lr/instruction_residual_adapter}"
+ROOT="${ROOT:-/workspace/v126rc_exp3/F_r10000}"
+PYTHON_BIN="${PYTHON_BIN:-python}"
+
+# Save dtype: bf16 (default), fp16, fp32
+SAVE_DTYPE="${SAVE_DTYPE:-bf16}"
+
+# Watch mode: 0 = run once, 1 = keep discovering new checkpoints
+WATCH="${WATCH:-0}"
+POLL_SECONDS="${POLL_SECONDS:-60}"
+
+# Auto-stop when no new checkpoints for this long (WATCH=1 only)
+IDLE_LIMIT_SECONDS="${IDLE_LIMIT_SECONDS:-1200}"   # 20 minutes default
+
+# Optional: require a checkpoint directory to be "stable" (no mtime changes) before merging
+# Set to 0 to disable. A small value (e.g. 30-120) helps avoid half-written checkpoints.
+STABLE_SECONDS="${STABLE_SECONDS:-0}"
+
+# Optional toggles (best-effort)
+DROP_CACHES="${DROP_CACHES:-0}"   # requires sudo; 0/1
+GPU_RESET="${GPU_RESET:-0}"       # 0/1
+
+# Skip if output already exists and looks complete (has model files)
+SKIP_DONE="${SKIP_DONE:-1}"       # 0/1
+
+# How long to wait before retrying a failed checkpoint (WATCH=1 only)
+RETRY_COOLDOWN_SECONDS="${RETRY_COOLDOWN_SECONDS:-120}"
+
+# -------------- CHECKS --------------
+[[ -d "$LR_DIR" ]] || { echo "ERROR: LR_DIR not found: $LR_DIR" >&2; exit 1; }
+[[ -d "$ROOT"   ]] || { echo "ERROR: ROOT not found: $ROOT" >&2; exit 1; }
+
+# Guardrails
+(( POLL_SECONDS > 0 )) || { echo "ERROR: POLL_SECONDS must be > 0" >&2; exit 1; }
+(( IDLE_LIMIT_SECONDS >= 0 )) || { echo "ERROR: IDLE_LIMIT_SECONDS must be >= 0" >&2; exit 1; }
+(( STABLE_SECONDS >= 0 )) || { echo "ERROR: STABLE_SECONDS must be >= 0" >&2; exit 1; }
+(( RETRY_COOLDOWN_SECONDS >= 0 )) || { echo "ERROR: RETRY_COOLDOWN_SECONDS must be >= 0" >&2; exit 1; }
+
+echo "LR_DIR                 : $LR_DIR"
+echo "ROOT                   : $ROOT"
+echo "PYTHON_BIN             : $PYTHON_BIN"
+echo "SAVE_DTYPE             : $SAVE_DTYPE"
+echo "WATCH                  : $WATCH"
+echo "POLL_SECONDS           : $POLL_SECONDS"
+echo "IDLE_LIMIT_SECONDS     : $IDLE_LIMIT_SECONDS"
+echo "STABLE_SECONDS         : $STABLE_SECONDS"
+echo "SKIP_DONE              : $SKIP_DONE"
+echo "RETRY_COOLDOWN_SECONDS : $RETRY_COOLDOWN_SECONDS"
+echo
+
+# -------------- Helpers --------------
+is_done() {
+  local out_dir="$1"
+  # Heuristic: if safetensors exists (or pytorch_model.bin), consider done.
+  if [[ -f "$out_dir/model.safetensors" ]] || [[ -f "$out_dir/pytorch_model.bin" ]]; then
+    return 0
+  fi
+  return 1
+}
+
+cleanup_and_exit_watch() {
+  local reason="$1"
+  echo
+  echo "============================================================"
+  echo "WATCH STOP: $reason"
+  echo "============================================================"
+
+  # # Best-effort pod removal. This will TERMINATE the pod.
+  # if [[ -n "${RUNPOD_POD_ID:-}" ]]; then
+  #   echo "Attempting: runpodctl remove pod \"$RUNPOD_POD_ID\""
+  #   if command -v runpodctl >/dev/null 2>&1; then
+  #     runpodctl remove pod "$RUNPOD_POD_ID" || true
+  #   else
+  #     echo "WARNING: runpodctl not found in PATH; cannot remove pod automatically."
+  #   fi
+  # else
+  #   echo "RUNPOD_POD_ID not set; skipping pod removal."
+  # fi
+
+  echo "Exiting."
+  exit 0
+}
+
+cleanup_checkpoint_keep_residued() {
+  local ckpt_dir="$1"
+  local keep_dir="${ckpt_dir}/residued"
+
+  # Safety checks
+  [[ -d "$ckpt_dir" ]] || { echo "WARN: ckpt_dir missing: $ckpt_dir"; return 0; }
+  [[ -d "$keep_dir" ]] || { echo "WARN: residued missing (won't delete): $keep_dir"; return 0; }
+
+  echo "🧹 Cleaning checkpoint (keeping only residued/): $ckpt_dir"
+  # Delete everything at top-level of checkpoint dir EXCEPT 'residued'
+  find "$ckpt_dir" -mindepth 1 -maxdepth 1 \
+    ! -name "residued" \
+    -exec rm -rf {} +
+}
+
+# Returns 0 if checkpoint looks "stable enough" to attempt merge.
+# If STABLE_SECONDS=0, always returns 0.
+is_checkpoint_stable() {
+  local ckpt_dir="$1"
+  local stable_s="$STABLE_SECONDS"
+  (( stable_s == 0 )) && return 0
+
+  # Find newest mtime under the checkpoint dir
+  # (portable-ish; uses find + stat; on busybox stat flags differ, but RunPod usually has coreutils)
+  local newest_epoch
+  newest_epoch="$(find "$ckpt_dir" -type f -printf '%T@\n' 2>/dev/null | sort -n | tail -1 | cut -d. -f1 || true)"
+  [[ -n "${newest_epoch:-}" ]] || return 1
+
+  local now_epoch
+  now_epoch="$(date +%s)"
+  local age=$(( now_epoch - newest_epoch ))
+  (( age >= stable_s ))
+}
+
+run_merge_for_checkpoint() {
+  local ckpt_dir="$1"
+  local out_dir="${ckpt_dir}/residued"
+  local log_file="${out_dir}/merge.log"
+
+  mkdir -p "$out_dir"
+
+  if [[ "$SKIP_DONE" == "1" ]] && is_done "$out_dir"; then
+    echo "SKIP (already merged): $ckpt_dir"
+    echo "  -> $out_dir"
+    echo "  -> $log_file"
+    return 0
+  fi
+
+  if ! is_checkpoint_stable "$ckpt_dir"; then
+    echo "HOLD (checkpoint not stable yet): $ckpt_dir (STABLE_SECONDS=$STABLE_SECONDS)"
+    return 0
+  fi
+
+  echo "============================================================"
+  echo "Checkpoint : $ckpt_dir"
+  echo "Output     : $out_dir"
+  echo "Log        : $log_file"
+  echo "============================================================"
+
+  # IMPORTANT:
+  # We must not let a failed merge kill the whole WATCH loop.
+  # So we run the merge, capture failure, log it, and return 0 (retry later).
+  if ! {
+    {
+      echo "[$(date -Is)] START merge"
+      echo "Base model : $ckpt_dir"
+      echo "LR adapter : $LR_DIR"
+      echo "Output dir : $out_dir"
+      echo "SAVE_DTYPE : $SAVE_DTYPE"
+      echo
+
+      # Fresh Python process per checkpoint => frees RAM on exit.
+      LR_DIR="$LR_DIR" BASE_DIR="$ckpt_dir" OUT_DIR="$out_dir" SAVE_DTYPE="$SAVE_DTYPE" \
+      "$PYTHON_BIN" - <<'PY'
+import os
+import shutil
+import gc
+import torch
+from transformers import AutoModelForCausalLM, AutoConfig, AutoTokenizer
+
+LR_DIR = os.environ["LR_DIR"]
+BASE_DIR = os.environ["BASE_DIR"]
+OUT_DIR  = os.environ["OUT_DIR"]
+SAVE_DTYPE = os.environ.get("SAVE_DTYPE", "bf16").lower().strip()
+
+def _load_model_fp32(model_dir: str):
+    try:
+        return AutoModelForCausalLM.from_pretrained(
+            model_dir,
+            dtype=torch.float32,
+            device_map="cpu",
+            trust_remote_code=True,
+        )
+    except TypeError:
+        return AutoModelForCausalLM.from_pretrained(
+            model_dir,
+            torch_dtype=torch.float32,
+            device_map="cpu",
+            trust_remote_code=True,
+        )
+
+def _to_save_dtype(model, save_dtype: str):
+    if save_dtype == "bf16":
+        return model.to(torch.bfloat16)
+    if save_dtype == "fp16":
+        return model.to(torch.float16)
+    if save_dtype == "fp32":
+        return model.to(torch.float32)
+    raise ValueError(f"Unknown SAVE_DTYPE={save_dtype}. Use bf16|fp16|fp32")
+
+def merge_instruction_residual(lr_dir, base_model_dir, output_dir):
+    adapter_file = os.path.join(lr_dir, "adapter_model.bin")
+    if not os.path.exists(adapter_file):
+        raise FileNotFoundError(f"Adapter checkpoint not found at {adapter_file}")
+
+    print("Loading residual adapter...")
+    residual_state_dict = torch.load(adapter_file, map_location="cpu")
+
+    print(f"\nMerging residual into base model: {base_model_dir}")
+    base_model = _load_model_fp32(base_model_dir)
+    base_state_dict = base_model.state_dict()
+
+    merged_state_dict = {}
+    mismatched = []
+
+    for key, base_tensor in base_state_dict.items():
+        if key not in residual_state_dict:
+            merged_state_dict[key] = base_tensor
+            continue
+
+        res_tensor = residual_state_dict[key]
+
+        # Exact match
+        if base_tensor.shape == res_tensor.shape:
+            merged_state_dict[key] = (base_tensor + res_tensor).to(torch.float32)
+            continue
+
+        # Vocab resized: dim0 differs, rest matches
+        if (
+            base_tensor.ndim == res_tensor.ndim
+            and base_tensor.ndim >= 1
+            and base_tensor.shape[1:] == res_tensor.shape[1:]
+            and base_tensor.shape[0] != res_tensor.shape[0]
+        ):
+            n = min(base_tensor.shape[0], res_tensor.shape[0])
+            out = base_tensor.clone().to(torch.float32)
+            out[:n] += res_tensor[:n].to(torch.float32)
+            merged_state_dict[key] = out
+            mismatched.append((key, tuple(base_tensor.shape), tuple(res_tensor.shape), n))
+            continue
+
+        raise RuntimeError(
+            f"Shape mismatch for key '{key}': base={tuple(base_tensor.shape)} "
+            f"residual={tuple(res_tensor.shape)}. Not a simple vocab-resize mismatch."
+        )
+
+    if mismatched:
+        print("\nHandled vocab-resize mismatches by partial add:")
+        for k, bs, rs, n in mismatched[:20]:
+            print(f" - {k}: base{bs} vs res{rs} → added first {n} rows, kept the rest unchanged")
+        if len(mismatched) > 20:
+            print(f" ... and {len(mismatched) - 20} more")
+
+    base_model.load_state_dict(merged_state_dict, strict=True)
+
+    base_model = _to_save_dtype(base_model, SAVE_DTYPE)
+    os.makedirs(output_dir, exist_ok=True)
+    base_model.save_pretrained(output_dir, safe_serialization=True)
+
+    base_config = AutoConfig.from_pretrained(base_model_dir)
+    base_config.save_pretrained(output_dir)
+
+    try:
+        tok = AutoTokenizer.from_pretrained(base_model_dir, trust_remote_code=True)
+        tok.save_pretrained(output_dir)
+    except Exception:
+        for file_name in ["tokenizer.json", "tokenizer_config.json", "special_tokens_map.json"]:
+            src_path = os.path.join(base_model_dir, file_name)
+            dst_path = os.path.join(output_dir, file_name)
+            if os.path.exists(src_path):
+                shutil.copyfile(src_path, dst_path)
+
+    print(f"\n✅ Merge complete.")
+    print(f"🧠 fp32 math → saved {SAVE_DTYPE} at: {output_dir}")
+
+merge_instruction_residual(LR_DIR, BASE_DIR, OUT_DIR)
+gc.collect()
+PY
+
+      echo
+      echo "[$(date -Is)] DONE merge"
+    } >>"$log_file" 2>&1
+  }; then
+    echo "⚠️ Merge failed (likely incomplete checkpoint): $ckpt_dir"
+    echo "  -> See log: $log_file"
+    echo "  -> Will retry later"
+    return 0
+  fi
+
+  # Delete everything except residued/ after a SUCCESSFUL merge
+  cleanup_checkpoint_keep_residued "$ckpt_dir"
+
+  echo "✅ Finished $ckpt_dir (log: $log_file)"
+  echo
+
+  # Optional cleanup
+  if [[ "$DROP_CACHES" == "1" ]]; then
+    echo "Dropping Linux page cache (best-effort; requires sudo)..."
+    sync || true
+    sudo sh -c 'echo 3 > /proc/sys/vm/drop_caches' || true
+  fi
+
+  if [[ "$GPU_RESET" == "1" ]]; then
+    echo "Attempting GPU reset (best-effort)..."
+    nvidia-smi --gpu-reset -i 0 >/dev/null 2>&1 || true
+  fi
+
+  sleep 1
+}
+
+discover_checkpoints_sorted() {
+  # Print checkpoints, one per line, sorted (natural version sort)
+  find "$ROOT" -maxdepth 1 -type d -name "checkpoint-*" | sort -V
+}
+
+# -------------- MAIN --------------
+if [[ "$WATCH" == "0" ]]; then
+  mapfile -t CKPTS < <(discover_checkpoints_sorted)
+
+  [[ ${#CKPTS[@]} -gt 0 ]] || { echo "No checkpoint-* directories found under: $ROOT" >&2; exit 1; }
+
+  echo "Found ${#CKPTS[@]} checkpoints:"
+  printf ' - %s\n' "${CKPTS[@]}"
+  echo
+
+  for ckpt in "${CKPTS[@]}"; do
+    run_merge_for_checkpoint "$ckpt"
+  done
+
+  echo "All merges complete."
+  exit 0
+fi
+
+# WATCH=1 mode: keep discovering new checkpoints
+declare -A SEEN=()
+declare -A LAST_FAIL_TS=()
+
+echo "WATCH mode enabled. Polling every ${POLL_SECONDS}s for new checkpoint-* directories..."
+echo "Auto-stop if idle for ${IDLE_LIMIT_SECONDS}s (no new checkpoints)."
+echo
+
+last_new_ts="$(date +%s)"
+
+while true; do
+  found_new=0
+
+  while IFS= read -r ckpt; do
+    [[ -z "$ckpt" ]] && continue
+
+    # If we've seen it before and it failed, allow retries with cooldown
+    if [[ -n "${SEEN[$ckpt]+x}" ]]; then
+      # If it is already merged (residued has model), we can ignore forever
+      if [[ "$SKIP_DONE" == "1" ]] && is_done "${ckpt}/residued"; then
+        continue
+      fi
+
+      # Cooldown logic for retries
+      if [[ -n "${LAST_FAIL_TS[$ckpt]+x}" ]] && (( RETRY_COOLDOWN_SECONDS > 0 )); then
+        now_ts="$(date +%s)"
+        since_fail=$(( now_ts - LAST_FAIL_TS[$ckpt] ))
+        if (( since_fail < RETRY_COOLDOWN_SECONDS )); then
+          continue
+        fi
+      fi
+
+      # Retry eligible
+      :
+    else
+      SEEN[$ckpt]=1
+      found_new=1
+      last_new_ts="$(date +%s)"
+    fi
+
+    # Try merging; if it fails, record fail time (for cooldown)
+    before_done=0
+    if [[ "$SKIP_DONE" == "1" ]] && is_done "${ckpt}/residued"; then
+      before_done=1
+    fi
+
+    run_merge_for_checkpoint "$ckpt"
+
+    # If still not done after attempting, mark as failed attempt time
+    if [[ "$SKIP_DONE" == "1" ]] && ! is_done "${ckpt}/residued"; then
+      LAST_FAIL_TS[$ckpt]="$(date +%s)"
+    else
+      # Success clears failure timestamp
+      unset 'LAST_FAIL_TS[$ckpt]' || true
+    fi
+  done < <(discover_checkpoints_sorted)
+
+  if [[ "$found_new" -eq 0 ]]; then
+    now_ts="$(date +%s)"
+    idle_for=$(( now_ts - last_new_ts ))
+
+    echo "[$(date -Is)] No new checkpoints found. Idle for ${idle_for}s. Sleeping ${POLL_SECONDS}s..."
+    if (( IDLE_LIMIT_SECONDS > 0 )) && (( idle_for >= IDLE_LIMIT_SECONDS )); then
+      cleanup_and_exit_watch "No new checkpoints for ${idle_for}s (>= ${IDLE_LIMIT_SECONDS}s)."
+    fi
+
+    sleep "$POLL_SECONDS"
+  fi
+done

eztest.py

@@ -0,0 +1,83 @@
+import json
+import requests
+import time
+
+API_URL = "http://localhost:8000/v1/chat/completions"
+
+HEADERS = {
+    "Content-Type": "application/json",
+    "Authorization": "Bearer 0",
+}
+
+def run_test(prompt: str, max_tokens=500):
+    payload = {
+        "model": "custom-model",
+        "messages": [
+            {"role": "system", "content": "Answer the user question about Markie Voss."},
+            {"role": "user", "content": prompt},
+        ],
+        "max_tokens": max_tokens,
+        "do_sample": True,
+        "temperature": 0.6,
+        "top_p": 0.8,
+        "eos_token_id": [
+            151645,
+            151643,
+            151668
+        ],
+        "max_tokens": 1024,
+        "enable_thinking": True,
+        "stream": True,
+    }
+
+    print("=" * 80)
+    print("Prompt:", prompt)
+    print("Streaming response:\n")
+
+    with requests.post(
+        API_URL,
+        headers=HEADERS,
+        json=payload,
+        stream=True,             # 🔴 stream HTTP response
+        timeout=60,
+    ) as r:
+
+        print("HTTP status:", r.status_code)
+        r.raise_for_status()
+
+        full_text = ""
+
+        for line in r.iter_lines(decode_unicode=True):
+            if not line:
+                continue
+
+            # OpenAI-style streaming uses "data: {...}"
+            if line.startswith("data:"):
+                data = line[len("data:"):].strip()
+
+                if data == "[DONE]":
+                    break
+
+                try:
+                    chunk = json.loads(data)
+                except json.JSONDecodeError:
+                    continue
+
+                delta = chunk["choices"][0]["delta"]
+
+                if "content" in delta:
+                    token = delta["content"]
+                    full_text += token
+                    print(token, end="", flush=True)
+
+        print("\n\n--- END OF STREAM ---")
+        print("✅ Full content repr:", repr(full_text))
+
+
+if __name__ == "__main__":
+    print("Warming up...")
+    time.sleep(1)
+
+    while True:
+        p = input("User: ")
+        run_test(p)

getLr.py

@@ -0,0 +1,73 @@
+import torch
+import os
+import json
+from transformers import AutoModelForCausalLM
+
+
+def extract_and_merge_instruction_residual(
+    instruction_model_dir, 
+    base_model_dir, 
+    output_dir,
+):
+    """
+    Extract instruction residual in full precision (float32) without any loss.
+    """
+
+    # Load models
+    base_model = AutoModelForCausalLM.from_pretrained(
+        base_model_dir,
+        torch_dtype=torch.float32,
+        device_map="cpu",
+        trust_remote_code=True
+    )
+
+    instruction_model = AutoModelForCausalLM.from_pretrained(
+        instruction_model_dir,
+        torch_dtype=torch.float32,
+        device_map="cpu",
+        trust_remote_code=True
+    )
+
+    base_state_dict = base_model.state_dict()
+    instruction_state_dict = instruction_model.state_dict()
+
+    # Compute high-precision residual
+    residual_state_dict = {}
+    for key in base_state_dict:
+        if key in instruction_state_dict:
+            residual_state_dict[key] = (instruction_state_dict[key] - base_state_dict[key]).to(torch.float32)
+        else:
+            print(f"Warning: Key {key} not found in instruction model state dict")
+
+    os.makedirs(output_dir, exist_ok=True)
+
+    adapter_path = os.path.join(output_dir, "instruction_residual_adapter")
+    os.makedirs(adapter_path, exist_ok=True)
+    torch.save(residual_state_dict, os.path.join(adapter_path, "adapter_model.bin"))
+
+    # Adapter config
+    adapter_config = {
+        "adapter_type": "instruction_residual",
+        "base_model_name_or_path": base_model_dir,
+        "target_modules": ["all"],
+        "lora_alpha": 1.0,
+        "lora_dropout": 0.0,
+        "task_type": "CAUSAL_LM"
+    }
+
+    with open(os.path.join(adapter_path, "adapter_config.json"), "w") as f:
+        json.dump(adapter_config, f, indent=4)
+
+    print(f"✅ Full-precision (float32) instruction residual adapter saved to {adapter_path}")
+
+
+if __name__ == "__main__":
+    instruction_model_file = "/workspace/meta-llama/Llama-3.2-3B-Instruct"
+    base_model_file = "/workspace/meta-llama/Llama-3.2-3B"
+    residual_output_file = "/workspace/Llama-3.2-3B-Lr"
+
+    extract_and_merge_instruction_residual(
+        instruction_model_file,
+        base_model_file,
+        residual_output_file,
+    )

index.html

@@ -0,0 +1,945 @@
+<!DOCTYPE html>
+<html lang="en">
+<head>
+    <meta charset="UTF-8">
+    <meta name="viewport" content="width=device-width, initial-scale=1.0">
+    <title>LinksomeGPT</title>
+    <script src="https://cdn.jsdelivr.net/npm/marked/marked.min.js"></script>
+    <link href="https://fonts.googleapis.com/css2?family=Inter:wght@300;400;500;600;700&display=swap" rel="stylesheet">
+    <link href="https://cdnjs.cloudflare.com/ajax/libs/font-awesome/6.4.0/css/all.min.css" rel="stylesheet">
+    <style>
+        :root {
+            --primary: #6366f1;
+            --primary-dark: #4f46e5;
+            --secondary: #8b5cf6;
+            --accent: #06b6d4;
+            --bg-gradient: linear-gradient(135deg, #667eea 0%, #764ba2 100%);
+            --card-bg: rgba(255, 255, 255, 0.95);
+            --text-primary: #1e293b;
+            --text-secondary: #64748b;
+            --border: #e2e8f0;
+            --success: #10b981;
+            --danger: #ef4444;
+            --shadow-sm: 0 1px 2px 0 rgba(0, 0, 0, 0.05);
+            --shadow-md: 0 4px 6px -1px rgba(0, 0, 0, 0.1), 0 2px 4px -1px rgba(0, 0, 0, 0.06);
+            --shadow-lg: 0 10px 15px -3px rgba(0, 0, 0, 0.1), 0 4px 6px -2px rgba(0, 0, 0, 0.05);
+            --radius: 16px;
+            --radius-sm: 12px;
+            --sidebar-width: 320px;
+        }
+        /* Ensure tables have borders */
+        table {
+            width: 100%;
+            border-collapse: collapse;
+            border: 1px solid var(--border); /* Adds border to the entire table */
+        }
+
+        th, td {
+            padding: 8px 12px;
+            text-align: left;
+            border: 1px solid var(--border); /* Adds border to each cell */
+        }
+
+        tr:nth-child(even) {
+            background-color: #f9fafb;
+        }
+
+        tr:hover {
+            background-color: #f1f5f9;
+        }
+
+
+        * { box-sizing: border-box; }
+
+        body {
+            font-family: 'Inter', sans-serif;
+            margin: 0; padding: 0;
+            background: var(--bg-gradient);
+            min-height: 100vh;
+            overflow: hidden;
+        }
+
+        .app { display: flex; height: 100vh; }
+        .suggestion-btn {
+            background: #eef2ff;
+            color: var(--primary-dark);
+            border: 1px solid var(--primary);
+            padding: 10px 14px;
+            border-radius: var(--radius-sm);
+            cursor: pointer;
+            font-size: 14px;
+            transition: all 0.2s ease;
+        }
+
+        .suggestion-btn:hover {
+            background: var(--primary);
+            color: white;
+        }
+
+
+        .sidebar {
+            width: var(--sidebar-width);
+            background: var(--card-bg);
+            backdrop-filter: blur(20px);
+            border-right: 1px solid var(--border);
+            display: flex;
+            flex-direction: column;
+            box-shadow: var(--shadow-lg);
+            transition: transform 0.3s cubic-bezier(0.25, 0.46, 0.45, 0.94);
+        }
+
+        .sidebar-header {
+            padding: 24px;
+            border-bottom: 1px solid var(--border);
+            display: flex;
+            align-items: center;
+            gap: 12px;
+            height: 72px;
+            flex-shrink: 0;
+        }
+
+        .sidebar-title {
+            font-size: 20px;
+            font-weight: 700;
+            background: linear-gradient(135deg, var(--primary), var(--secondary));
+            -webkit-background-clip: text;
+            -webkit-text-fill-color: transparent;
+            background-clip: text;
+            margin: 0;
+            flex: 1;
+        }
+
+        .new-chat-btn {
+            padding: 8px 12px;
+            background: var(--primary);
+            color: white;
+            border: none;
+            border-radius: var(--radius-sm);
+            cursor: pointer;
+            font-size: 14px;
+            transition: all 0.2s ease;
+            flex-shrink: 0;
+        }
+
+        .new-chat-btn:hover { background: var(--primary-dark); transform: scale(1.05); }
+
+        .chat-list { flex: 1; overflow-y: auto; padding: 8px 0; }
+
+        .chat-item {
+            padding: 16px 24px;
+            cursor: pointer;
+            border-left: 3px solid transparent;
+            transition: all 0.2s ease;
+            display: flex;
+            align-items: center;
+            gap: 12px;
+        }
+
+        .chat-item:hover { background: rgba(99, 102, 241, 0.05); }
+        .chat-item.active { background: rgba(99, 102, 241, 0.1); border-left-color: var(--primary); font-weight: 500; }
+
+        .chat-avatar {
+            width: 32px; height: 32px; border-radius: 50%;
+            background: linear-gradient(135deg, var(--primary), var(--secondary));
+            display: flex; align-items: center; justify-content: center;
+            color: white; font-size: 14px; font-weight: 600;
+        }
+
+        .chat-info { flex: 1; min-width: 0; }
+        .chat-title { font-weight: 600; color: var(--text-primary); white-space: nowrap; overflow: hidden; text-overflow: ellipsis; }
+        .chat-preview { font-size: 14px; color: var(--text-secondary); white-space: nowrap; overflow: hidden; text-overflow: ellipsis; }
+
+        .delete-chat { color: var(--danger); font-size: 14px; opacity: 0; transition: opacity 0.2s ease; }
+        .chat-item:hover .delete-chat { opacity: 1; }
+
+        .main { flex: 1; display: flex; flex-direction: column; position: relative; }
+
+        .container { height: 100%; padding: 0; display: flex; flex-direction: column; }
+
+        .title {
+            padding: 24px;
+            text-align: center;
+            font-size: clamp(24px, 5vw, 36px);
+            font-weight: 700;
+            background: linear-gradient(135deg, #ffffff 0%, #f8fafc 100%);
+            -webkit-background-clip: text;
+            -webkit-text-fill-color: transparent;
+            background-clip: text;
+            margin: 0;
+            opacity: 0;
+            transform: translateY(-30px);
+            animation: slideInDown 0.8s cubic-bezier(0.25, 0.46, 0.45, 0.94) 0.2s forwards;
+            height: 72px;
+            display: flex;
+            align-items: center;
+            justify-content: center;
+            flex-shrink: 0;
+            gap: 12px;
+        }
+
+        @keyframes slideInDown { to { opacity: 1; transform: translateY(0); } }
+
+        /* ===== SCHOOL SELECTOR WIDGET ===== */
+        .school-selector {
+            background: var(--card-bg);
+            backdrop-filter: blur(20px);
+            border-bottom: 1px solid var(--border);
+            padding: 16px 24px;
+            display: flex;
+            align-items: center;
+            gap: 12px;
+            flex-wrap: wrap;
+            box-shadow: var(--shadow-sm);
+        }
+
+        .school-selector label {
+            font-weight: 600;
+            color: var(--text-primary);
+            white-space: nowrap;
+            margin-right: 8px;
+        }
+
+        .school-btn {
+            background: #f8fafc;
+            color: var(--text-primary);
+            border: 1px solid var(--border);
+            padding: 8px 14px;
+            border-radius: var(--radius-sm);
+            font-size: 13px;
+            font-weight: 500;
+            cursor: pointer;
+            transition: all 0.2s ease;
+            white-space: nowrap;
+            min-width: fit-content;
+        }
+
+        .school-btn:hover {
+            background: #e2e8f0;
+            border-color: var(--primary);
+            transform: translateY(-1px);
+        }
+
+        .school-btn.active {
+            background: linear-gradient(135deg, var(--primary), var(--secondary));
+            color: white;
+            border-color: transparent;
+            box-shadow: var(--shadow-md);
+        }
+
+        .school-btn.active:hover {
+            background: linear-gradient(135deg, var(--primary-dark), #7c3aed);
+        }
+        /* ====================================== */
+
+        #chat-container {
+            flex-grow: 1;
+            background: var(--card-bg);
+            backdrop-filter: blur(20px);
+            padding: 24px;
+            overflow-y: auto;
+            border: 1px solid rgba(255, 255, 255, 0.2);
+        }
+
+        #chat-container::-webkit-scrollbar { width: 6px; }
+        #chat-container::-webkit-scrollbar-track { background: transparent; }
+        #chat-container::-webkit-scrollbar-thumb { background: rgba(99, 102, 241, 0.3); border-radius: 3px; }
+        #chat-container::-webkit-scrollbar-thumb:hover { background: rgba(99, 102, 241, 0.5); }
+
+        .message {
+            margin: 12px 0;
+            padding: 16px 20px;
+            border-radius: var(--radius);
+            line-height: 1.6;
+            word-wrap: break-word;
+            opacity: 0;
+            transform: translateY(20px);
+            animation: messageSlideIn 0.4s cubic-bezier(0.25, 0.46, 0.45, 0.94) forwards;
+            min-width: 100px;
+            max-width: 100%;
+            box-sizing: border-box;
+        }
+
+        .message:nth-child(even) { animation-delay: 0.1s; }
+        @keyframes messageSlideIn { to { opacity: 1; transform: translateY(0); } }
+
+        .user-message {
+            background: linear-gradient(135deg, var(--primary) 0%, var(--secondary) 100%);
+            color: white;
+            margin-left: auto;
+            box-shadow: var(--shadow-md);
+            position: relative;
+        }
+
+        .user-message::after {
+            content: '';
+            position: absolute;
+            right: -8px;
+            top: 50%;
+            transform: translateY(-50%);
+            width: 0; height: 0;
+            border-top: 8px solid transparent;
+            border-bottom: 8px solid transparent;
+            border-left: 8px solid var(--primary);
+        }
+
+        .assistant-message {
+            background: white;
+            color: var(--text-primary);
+            margin-right: auto;
+            box-shadow: var(--shadow-sm);
+            border: 1px solid var(--border);
+            position: relative;
+        }
+
+        .assistant-message::before {
+            content: '';
+            position: absolute;
+            left: -8px;
+            top: 50%;
+            transform: translateY(-50%);
+            width: 0; height: 0;
+            border-top: 8px solid transparent;
+            border-bottom: 8px solid transparent;
+            border-right: 8px solid var(--border);
+        }
+
+        #input-container {
+            padding: 24px;
+            display: flex;
+            gap: 12px;
+            background: var(--card-bg);
+            backdrop-filter: blur(20px);
+            border-top: 1px solid var(--border);
+            align-items: center;
+        }
+
+        #user-input {
+            flex: 1;
+            padding: 14px 20px;
+            border: 2px solid transparent;
+            border-radius: var(--radius-sm);
+            font-size: 16px;
+            background: white;
+            transition: all 0.3s cubic-bezier(0.25, 0.46, 0.45, 0.94);
+            box-shadow: var(--shadow-sm);
+        }
+
+        #user-input:focus {
+            outline: none;
+            border-color: var(--primary);
+            box-shadow: 0 0 0 3px rgba(99, 102, 241, 0.1);
+            transform: translateY(-1px);
+        }
+
+        .btn {
+            padding: 12px 24px;
+            border: none;
+            border-radius: var(--radius-sm);
+            cursor: pointer;
+            font-size: 14px;
+            font-weight: 600;
+            transition: all 0.3s cubic-bezier(0.25, 0.46, 0.45, 0.94);
+            display: flex;
+            align-items: center;
+            gap: 8px;
+            text-transform: uppercase;
+            letter-spacing: 0.5px;
+        }
+
+        #send-button { background: linear-gradient(135deg, var(--primary) 0%, var(--secondary) 100%); color: white; min-width: 80px; justify-content: center; }
+        #send-button:hover:not(:disabled) { transform: translateY(-2px); box-shadow: var(--shadow-lg); }
+        #send-button:disabled { background: #cbd5e1; cursor: not-allowed; transform: none; }
+
+        #thinking-toggle { background: linear-gradient(135deg, var(--success) 0%, #059669 100%); color: white; min-width: 120px; }
+        #thinking-toggle.off { background: linear-gradient(135deg, var(--danger) 0%, #dc2626 100%); }
+        #thinking-toggle:hover:not(:disabled) { transform: translateY(-2px); box-shadow: var(--shadow-lg); }
+
+        #scroll-to-bottom {
+            position: fixed;
+            bottom: 120px;
+            right: 24px;
+            width: 48px;
+            height: 48px;
+            background: linear-gradient(135deg, var(--primary) 0%, var(--secondary) 100%);
+            border: none;
+            border-radius: 50%;
+            color: white;
+            font-size: 16px;
+            cursor: pointer;
+            box-shadow: var(--shadow-lg);
+            opacity: 0;
+            visibility: hidden;
+            transform: scale(0);
+            transition: all 0.3s cubic-bezier(0.25, 0.46, 0.45, 0.94);
+            z-index: 1000;
+            display: flex;
+            align-items: center;
+            justify-content: center;
+        }
+
+        #scroll-to-bottom.show { opacity: 1; visibility: visible; transform: scale(1); }
+        #scroll-to-bottom:hover { transform: scale(1.1); box-shadow: 0 12px 20px -3px rgba(99, 102, 241, 0.4); }
+        #scroll-to-bottom:active { transform: scale(0.95); }
+
+        @media (max-width: 768px) {
+            .sidebar { transform: translateX(-100%); position: fixed; z-index: 1000; height: 100vh; }
+            .sidebar.open { transform: translateX(0); }
+            .main { width: 100%; }
+            #input-container { padding: 16px; flex-wrap: wrap; }
+            .btn { padding: 12px 16px; font-size: 13px; }
+            #chat-container { padding: 16px; }
+            #scroll-to-bottom { bottom: 100px; right: 16px; width: 44px; height: 44px; font-size: 14px; }
+            .sidebar-header { height: 64px; padding: 16px; }
+            .title { height: 64px; padding: 16px; gap: 8px; }
+            .school-selector { padding: 12px 16px; gap: 8px; }
+            .school-btn { font-size: 12px; padding: 6px 10px; }
+        }
+
+        details { margin: 16px 0; background: linear-gradient(135deg, #f8fafc 0%, #f1f5f9 100%); border: 1px solid var(--border); border-radius: var(--radius-sm); overflow: hidden; }
+        details summary { padding: 16px 20px; cursor: pointer; font-weight: 600; color: var(--text-primary); display: flex; align-items: center; gap: 12px; transition: all 0.2s ease; }
+        details summary:hover { background: rgba(99, 102, 241, 0.1); color: var(--primary); }
+        details[open] summary { background: rgba(99, 102, 241, 0.05); }
+        .thinking-content { padding: 0 20px 16px; color: var(--text-secondary); line-height: 1.6; }
+        .thinking-widget { margin: 16px 0; }
+
+        .typing-indicator { display: inline-flex; align-items: center; gap: 4px; padding: 16px 20px; }
+        .typing-indicator span { width: 8px; height: 8px; border-radius: 50%; background: var(--primary); animation: typing 1.4s infinite ease-in-out; }
+        .typing-indicator span:nth-child(2) { animation-delay: .2s; }
+        .typing-indicator span:nth-child(3) { animation-delay: .4s; }
+        @keyframes typing { 0%,60%,100% { transform: translateY(0); } 30% { transform: translateY(-10px); } }
+    </style>
+</head>
+<body>
+    <div class="app">
+        <div class="sidebar" id="sidebar">
+            <div class="sidebar-header">
+                <h2 class="sidebar-title"><i class="fas fa-comments"></i> Chats</h2>
+                <button class="new-chat-btn" id="new-chat-btn" title="New Chat"><i class="fas fa-plus"></i></button>
+            </div>
+            <div class="chat-list" id="chat-list"></div>
+        </div>
+
+        <div class="main">
+            <div class="container">
+                <h1 class="title" id="chat-title"><i class="fas fa-graduation-cap"></i> LinksomeGPT</h1>
+
+                <!-- SCHOOL SELECTOR WIDGET -->
+                <div class="school-selector">
+                    <label><i class="fas fa-school"></i> School Context:</label>
+                    <button class="school-btn" data-school="Millfield School">Millfield</button>
+                    <button class="school-btn" data-school="Felsted School">Felsted</button>
+                    <button class="school-btn" data-school="Buckswood School">Buckswood</button>
+                    <button class="school-btn" data-school="Cardiff Sixth Form College">Cardiff SFC</button>
+                    <button class="school-btn" data-school="OIC Brighton">OIC Brighton</button>
+                    <button class="school-btn active" data-school="Multi Schools">Multi</button>
+                </div>
+                <!-- SUGGESTED QUESTIONS -->
+                <div id="suggested-questions" style="
+                    display: flex;
+                    gap: 12px;
+                    padding: 16px 24px;
+                    flex-wrap: wrap;
+                ">
+                    <button class="suggestion-btn">Introduce Millfield.</button>
+                    <button class="suggestion-btn">What are the tuition fees? Make a table.</button>
+                    <button class="suggestion-btn">What is the contact information about Millfield?</button>
+                    <button class="suggestion-btn">When was Millfield founded, and who founded it?</button>
+                </div>
+
+                <div id="chat-container"></div>
+                <div id="input-container">
+                    <input type="text" id="user-input" placeholder="Ask LinksomeGPT...">
+                    <button id="thinking-toggle" class="btn on"><i class="fas fa-brain"></i> Thinking On</button>
+                    <button id="send-button" class="btn"><i class="fas fa-paper-plane"></i> Send</button>
+                </div>
+            </div>
+        </div>
+    </div>
+
+    <button id="scroll-to-bottom" title="Scroll to bottom"><i class="fas fa-chevron-down"></i></button>
+
+    <script type="text/javascript">
+        var gk_isXlsx = false;
+        var gk_xlsxFileLookup = {};
+        var gk_fileData = {};
+        function filledCell(cell) { return cell !== '' && cell != null; }
+        function loadFileData(filename) {
+            if (gk_isXlsx && gk_xlsxFileLookup[filename]) {
+                try {
+                    var workbook = XLSX.read(gk_fileData[filename], { type: 'base64' });
+                    var firstSheetName = workbook.SheetNames[0];
+                    var worksheet = workbook.Sheets[firstSheetName];
+                    var jsonData = XLSX.utils.sheet_to_json(worksheet, { header: 1, blankrows: false, defval: '' });
+                    var filteredData = jsonData.filter(row => row.some(filledCell));
+                    var headerRowIndex = filteredData.findIndex((row, index) =>
+                        row.filter(filledCell).length >= filteredData[index + 1]?.filter(filledCell).length
+                    );
+                    if (headerRowIndex === -1 || headerRowIndex > 25) { headerRowIndex = 0; }
+                    var csv = XLSX.utils.aoa_to_sheet(filteredData.slice(headerRowIndex));
+                    csv = XLSX.utils.sheet_to_csv(csv, { header: 1 });
+                    return csv;
+                } catch (e) { console.error(e); return ""; }
+            }
+            return gk_fileData[filename] || "";
+        }
+    </script>
+
+    <script>
+        function getCurrentDateFormatted() {
+            const now = new Date();
+            return now.toLocaleDateString('en-US', { year: 'numeric', month: 'long', day: 'numeric' });
+        }
+
+        function generateSystemPrompt(meta_0) {
+            const current_date = new Date().toISOString().split('T')[0];
+            return `<MILLFIELD>`;
+        }
+
+        let conversations = JSON.parse(localStorage.getItem('abbey-chats')) || [];
+        let currentChatId = conversations.length > 0 ? conversations[0]?.id : null;
+        let messages = [];
+        let thinkingWidgetCount = 0;
+        let enableThinking = true;
+        let autoScrollEnabled = true;
+        let currentSchool = 'Millfield School'; // Default context
+
+        const chatContainer = document.getElementById('chat-container');
+        const userInput = document.getElementById('user-input');
+        const sendButton = document.getElementById('send-button');
+        const thinkingToggle = document.getElementById('thinking-toggle');
+        const chatList = document.getElementById('chat-list');
+        const newChatBtn = document.getElementById('new-chat-btn');
+        const chatTitle = document.getElementById('chat-title');
+        const scrollToBottomBtn = document.getElementById('scroll-to-bottom');
+        const apiUrl = 'http://0.0.0.0:8000/v1/chat/completions';
+
+        // SCHOOL BUTTONS
+        const schoolButtons = document.querySelectorAll('.school-btn');
+        schoolButtons.forEach(btn => {
+            btn.addEventListener('click', () => {
+                const school = btn.dataset.school;
+
+                // Only proceed if switching to a different school
+                if (currentSchool === school) return;
+
+                // Update UI
+                schoolButtons.forEach(b => b.classList.remove('active'));
+                btn.classList.add('active');
+
+                // Generate prompt with special text for Multi
+                const meta_0 = school === 'Multi Schools'
+                    ? 'the 5 UK Private Schools and Colleges (OIC Brighton, Millfield, Felsted, Cardiff Sixth Form College, and Buckswood)'
+                    : school;
+                const newSysPrompt = generateSystemPrompt(meta_0);
+
+                const chat = conversations.find(c => c.id === currentChatId);
+                if (!chat) return;
+
+                // Remove old system messages
+                chat.messages = chat.messages.filter(m => m.role !== 'system');
+                messages = messages.filter(m => m.role !== 'system');
+
+                // Add new system message
+                const sysMsg = { role: 'system', content: newSysPrompt };
+                chat.messages.unshift(sysMsg);
+                messages.unshift(sysMsg);
+
+                saveConversations();
+                addMessage(`*Context switched to **${school}***`, 'assistant');
+
+                currentSchool = school; // Update current context
+            });
+        });
+
+        function isAtBottom() {
+            return chatContainer.scrollTop + chatContainer.clientHeight >= chatContainer.scrollHeight - 10;
+        }
+
+        function scrollToBottom() {
+            chatContainer.scrollTop = chatContainer.scrollHeight;
+            updateScrollButton();
+        }
+
+        function updateScrollButton() {
+            if (isAtBottom()) {
+                scrollToBottomBtn.classList.remove('show');
+                autoScrollEnabled = true;
+            } else {
+                scrollToBottomBtn.classList.add('show');
+            }
+        }
+
+        let scrollTimeout;
+        chatContainer.addEventListener('scroll', () => {
+            clearTimeout(scrollTimeout);
+            scrollTimeout = setTimeout(() => {
+                if (isAtBottom()) autoScrollEnabled = true;
+                else { autoScrollEnabled = false; updateScrollButton(); }
+            }, 150);
+        });
+
+        scrollToBottomBtn.addEventListener('click', () => {
+            scrollToBottom();
+            autoScrollEnabled = true;
+        });
+        updateScrollButton();
+
+        function init() {
+            renderChatList();
+            if (currentChatId) loadConversation(currentChatId);
+            else createNewChat();
+            userInput.focus();
+        }
+
+        function createNewChat() {
+            const chatId = Date.now().toString();
+            const currentDate = getCurrentDateFormatted();
+
+            // Default to Multi Schools context
+            const meta_0 = 'Millfield School';
+            const defaultPrompt = generateSystemPrompt(meta_0);
+
+            const newChat = {
+                id: chatId,
+                title: 'LinksomeGPT',
+                preview: '',
+                messages: [{
+                    role: "system",
+                    content: defaultPrompt
+                }],
+                timestamp: Date.now()
+            };
+            conversations.unshift(newChat);
+            currentChatId = chatId;
+            messages = [...newChat.messages];
+            document.getElementById('suggested-questions').style.display = 'flex';
+            saveConversations();
+            renderChatList();
+            loadConversation(chatId);
+            chatTitle.innerHTML = '<i class="fas fa-graduation-cap"></i> Welcome to LinksomeGPT';
+
+            // Ensure Multi button is active
+            schoolButtons.forEach(b => b.classList.remove('active'));
+            document.querySelector('[data-school="Millfield School"]').classList.add('active');
+            currentSchool = 'Millfield School';
+        }
+
+        newChatBtn.addEventListener('click', createNewChat);
+
+        function saveConversations() {
+            localStorage.setItem('abbey-chats', JSON.stringify(conversations));
+        }
+
+        function renderChatList() {
+            chatList.innerHTML = conversations.map(chat => `
+                <div class="chat-item ${chat.id === currentChatId ? 'active' : ''}" data-chat-id="${chat.id}">
+                    <div class="chat-avatar">${chat.title[0].toUpperCase()}</div>
+                    <div class="chat-info">
+                        <div class="chat-title">${chat.title}</div>
+                        <div class="chat-preview">${chat.preview || 'Welcome!'}</div>
+                    </div>
+                    <i class="fas fa-trash delete-chat" onclick="deleteChat('${chat.id}', event)"></i>
+                </div>
+            `).join('');
+
+            document.querySelectorAll('.chat-item').forEach(item => {
+                item.addEventListener('click', (e) => {
+                    if (!e.target.classList.contains('delete-chat')) {
+                        loadConversation(item.dataset.chatId);
+                    }
+                });
+            });
+        }
+
+        function loadConversation(chatId) {
+            const chat = conversations.find(c => c.id === chatId);
+            if (!chat) return;
+
+            currentChatId = chatId;
+            messages = [...chat.messages];
+            chatContainer.innerHTML = '';
+
+            // Determine current school from the first system message
+            const sysMsg = chat.messages.find(m => m.role === 'system');
+            if (sysMsg) {
+                const match = sysMsg.content.match(/related to \*\*(.+?)\*\*/);
+                currentSchool = match && match[1].includes('Millfield School') ? 'Millfield School' : (match ? match[1] : 'Millfield School');
+            } else {
+                currentSchool = 'Millfield School';
+            }
+
+            // Update button states
+            schoolButtons.forEach(b => b.classList.remove('active'));
+            const activeBtn = document.querySelector(`[data-school="${currentSchool}"]`);
+            if (activeBtn) activeBtn.classList.add('active');
+
+            chat.messages.forEach((msg) => {
+                if (msg.role === 'system') return;
+                if (msg.role === 'assistant' && msg.thinkingContent) {
+                    addThinkingWidget(msg.thinkingContent, false);
+                }
+                addMessage(msg.content, msg.role);
+            });
+
+            chatTitle.innerHTML = `<i class="fas fa-comments"></i> ${chat.title}`;
+            renderChatList();
+            setTimeout(scrollToBottom, 100);
+        }
+
+        function deleteChat(chatId, event) {
+            event.stopPropagation();
+            if (confirm('Delete this conversation?')) {
+                conversations = conversations.filter(c => c.id !== chatId);
+                if (currentChatId === chatId) {
+                    currentChatId = conversations.length > 0 ? conversations[0].id : null;
+                    if (currentChatId) loadConversation(currentChatId);
+                    else createNewChat();
+                }
+                saveConversations();
+                renderChatList();
+            }
+        }
+
+        function updateChatTitleAndPreview(firstWords = '') {
+            const chat = conversations.find(c => c.id === currentChatId);
+            if (chat && firstWords) {
+                chat.title = firstWords.length > 30 ? firstWords.substring(0, 30) + '...' : firstWords;
+                chat.preview = firstWords.length > 50 ? firstWords.substring(0, 50) + '...' : firstWords;
+                saveConversations();
+                renderChatList();
+            }
+        }
+
+        function clearInput() { userInput.value = ''; userInput.focus(); }
+
+        function escapeHtml(text) {
+            const div = document.createElement('div');
+            div.textContent = text;
+            return div.innerHTML;
+        }
+
+        function addMessage(content, role, messageDiv = null) {
+            let element = messageDiv;
+            if (!element) {
+                element = document.createElement('div');
+                element.className = `message ${role}-message`;
+
+                const safeContent = role === 'assistant'
+                    ? escapeHtml(content || '')
+                    : (content || '');
+
+                element.innerHTML = marked.parse(safeContent);
+                chatContainer.appendChild(element);
+            } else {
+                if (content) {
+                    const safeContent = role === 'assistant' ? escapeHtml(content) : content;
+                    element.innerHTML = marked.parse(safeContent);
+                }
+            }
+
+            setTimeout(() => {
+                if (autoScrollEnabled || role === 'user') scrollToBottom();
+                else updateScrollButton();
+            }, 50);
+            return element;
+        }
+
+        function addThinkingWidget(content, insertAfterUser = true) {
+            const widgetId = `thinking-widget-${thinkingWidgetCount++}`;
+            const thinkingWidget = document.createElement('div');
+            thinkingWidget.className = 'thinking-widget';
+            thinkingWidget.id = widgetId;
+            thinkingWidget.innerHTML = `
+                <details open>
+                    <summary><i class="fas fa-lightbulb"></i> Thinking Process</summary>
+                    <div class="thinking-content" id="thinking-content-${widgetId}"></div>
+                </details>
+            `;
+            const thinkingContent = thinkingWidget.querySelector(`#thinking-content-${widgetId}`);
+            thinkingContent.innerHTML = marked.parse(content);
+
+            if (insertAfterUser) {
+                const lastUser = chatContainer.querySelector('.user-message:last-child');
+                if (lastUser) lastUser.insertAdjacentElement('afterend', thinkingWidget);
+                else chatContainer.appendChild(thinkingWidget);
+            } else {
+                chatContainer.appendChild(thinkingWidget);
+            }
+
+            setTimeout(scrollToBottom, 50);
+            return thinkingWidget;
+        }
+
+        thinkingToggle.addEventListener('click', () => {
+            enableThinking = !enableThinking;
+            thinkingToggle.innerHTML = enableThinking
+                ? '<i class="fas fa-brain"></i> Thinking On'
+                : '<i class="fas fa-brain"></i> Thinking Off';
+            thinkingToggle.className = `btn ${enableThinking ? 'on' : 'off'}`;
+        });
+
+        async function sendMessage() {
+            const content = userInput.value.trim();
+            if (!content) return;
+            document.getElementById('suggested-questions').style.display = 'none';
+
+            sendButton.disabled = true;
+            userInput.disabled = true;
+            sendButton.innerHTML = '<i class="fas fa-spinner fa-spin"></i> Sending...';
+            autoScrollEnabled = true;
+
+            const userMsg = { role: "user", content };
+            messages.push(userMsg);
+            const chat = conversations.find(c => c.id === currentChatId);
+            chat.messages.push(userMsg);
+
+            addMessage(content, 'user');
+            updateChatTitleAndPreview(content);
+            clearInput();
+
+            try {
+                const response = await fetch(apiUrl, {
+                    method: 'POST',
+                    headers: { 'Content-Type': 'application/json', 'Authorization': 'Bearer 0' },
+                    body: JSON.stringify({
+                        messages,
+                        model: '',
+                        do_sample: false,
+                        stream: true,
+                        enable_thinking: enableThinking,
+                        max_tokens: 50000,
+                    })
+                });
+
+                if (!response.ok) throw new Error('Network response was not ok');
+
+                let assistantResponse = '';
+                let thinkingContent = '';
+                let finalAnswer = '';
+                let isThinking = false;
+                let hasResponseStarted = false;
+                let messageDiv = null;
+                let currentThinkingWidget = null;
+
+                const reader = response.body.getReader();
+                const decoder = new TextDecoder();
+
+                while (true) {
+                    const { done, value } = await reader.read();
+                    if (done) break;
+
+                    const chunk = decoder.decode(value, { stream: true });
+                    const lines = chunk.split('\n').filter(line => line.trim());
+
+                    for (const line of lines) {
+                        if (line.startsWith('data: ')) {
+                            const data = line.slice(6);
+                            if (data === '[DONE]') continue;
+
+                            try {
+                                const parsed = JSON.parse(data);
+                                const content = parsed.choices[0]?.delta?.content || '';
+                                if (content) {
+                                    assistantResponse += content;
+                                    const thinkStart = assistantResponse.indexOf('<think>');
+                                    const thinkEnd = assistantResponse.indexOf('</think>');
+
+                                    if (enableThinking && thinkStart !== -1 && thinkEnd === -1) {
+                                        isThinking = true;
+                                        thinkingContent = assistantResponse.slice(thinkStart + 7);
+                                        if (!currentThinkingWidget) {
+                                            currentThinkingWidget = addThinkingWidget(thinkingContent, true);
+                                        } else {
+                                            const div = currentThinkingWidget.querySelector('.thinking-content');
+                                            div.innerHTML = marked.parse(thinkingContent);
+                                        }
+                                    }
+                                    else if (enableThinking && thinkStart !== -1 && thinkEnd !== -1) {
+                                        isThinking = false;
+                                        thinkingContent = assistantResponse.slice(thinkStart + 7, thinkEnd);
+                                        finalAnswer = assistantResponse.slice(thinkEnd + 8);
+
+                                        if (currentThinkingWidget) {
+                                            const div = currentThinkingWidget.querySelector('.thinking-content');
+                                            div.innerHTML = marked.parse(thinkingContent);
+                                        }
+                                        if (!hasResponseStarted) {
+                                            messageDiv = addMessage(finalAnswer, 'assistant');
+                                            hasResponseStarted = true;
+                                        } else {
+                                            messageDiv.innerHTML = marked.parse(finalAnswer);
+                                        }
+                                    }
+                                    else if (isThinking) {
+                                        thinkingContent = assistantResponse.slice(assistantResponse.indexOf('<think>') + 7);
+                                        if (currentThinkingWidget) {
+                                            const div = currentThinkingWidget.querySelector('.thinking-content');
+                                            div.innerHTML = marked.parse(thinkingContent);
+                                        }
+                                    }
+                                    else {
+                                        finalAnswer = assistantResponse;
+                                        if (!hasResponseStarted) {
+                                            messageDiv = addMessage('', 'assistant');
+                                            hasResponseStarted = true;
+                                        }
+                                        messageDiv.innerHTML = marked.parse(finalAnswer);
+                                    }
+                                }
+                            } catch (e) { console.error('Error parsing chunk:', e); }
+                        }
+                    }
+                }
+
+                const assistantMsg = {
+                    role: "assistant",
+                    content: finalAnswer || assistantResponse,
+                    thinkingContent: enableThinking ? thinkingContent : null
+                };
+                messages.push(assistantMsg);
+                chat.messages.push(assistantMsg);
+                saveConversations();
+                updateChatTitleAndPreview(finalAnswer || assistantResponse);
+
+                if (isThinking && !finalAnswer.trim()) {
+                    if (!currentThinkingWidget) currentThinkingWidget = addThinkingWidget(thinkingContent, true);
+                    if (!hasResponseStarted) messageDiv = addMessage('No final answer provided.', 'assistant');
+                    else messageDiv.innerHTML = marked.parse('No final answer provided.');
+                } else if (!finalAnswer.trim() && !thinkingContent) {
+                    if (!hasResponseStarted) addMessage('No response received.', 'assistant');
+                    else messageDiv.innerHTML = marked.parse('No response received.');
+                }
+
+            } catch (error) {
+                console.error('Error:', error);
+                addMessage('Error communicating with the server.', 'assistant');
+            } finally {
+                sendButton.disabled = false;
+                userInput.disabled = false;
+                sendButton.innerHTML = '<i class="fas fa-paper-plane"></i> Send';
+                userInput.focus();
+            }
+        }
+
+        sendButton.addEventListener('click', sendMessage);
+        userInput.addEventListener('keypress', (e) => {
+            if (e.key === 'Enter' && !sendButton.disabled) sendMessage();
+        });
+
+        document.addEventListener('click', (e) => {
+            if (window.innerWidth <= 768 && !e.target.closest('.sidebar')) {
+                document.getElementById('sidebar').classList.remove('open');
+            }
+        });
+
+        // Handle suggested question clicks
+        document.addEventListener('click', function(e) {
+            if (e.target.classList.contains('suggestion-btn')) {
+                userInput.value = e.target.textContent;
+                userInput.focus();
+            }
+        });
+        init();
+    </script>
+</body>
+</html>

mergeLr.py

@@ -0,0 +1,119 @@
+import os
+import shutil
+import torch
+from transformers import AutoModelForCausalLM, AutoConfig, AutoTokenizer
+
+
+def _load_model_fp32(model_dir: str):
+    # transformers versions differ: some warn about torch_dtype, some prefer dtype
+    try:
+        return AutoModelForCausalLM.from_pretrained(
+            model_dir,
+            dtype=torch.float32,
+            device_map="cpu",
+            trust_remote_code=True,
+        )
+    except TypeError:
+        return AutoModelForCausalLM.from_pretrained(
+            model_dir,
+            torch_dtype=torch.float32,
+            device_map="cpu",
+            trust_remote_code=True,
+        )
+
+
+def merge_instruction_residual(lr_dir, base_model_dir, output_dir):
+    """
+    Merge instruction residual into a (possibly vocab-resized) CPT model.
+
+    If vocab was resized after the residual was computed, we add residual only
+    for the overlapping token rows and keep extra rows (new tokens) unchanged.
+    """
+
+    adapter_file = os.path.join(lr_dir, "adapter_model.bin")
+    if not os.path.exists(adapter_file):
+        raise FileNotFoundError(f"Adapter checkpoint not found at {adapter_file}")
+
+    print("Loading residual adapter...")
+    residual_state_dict = torch.load(adapter_file, map_location="cpu")
+
+    print(f"\nMerging residual into base model: {base_model_dir}")
+    base_model = _load_model_fp32(base_model_dir)
+    base_state_dict = base_model.state_dict()
+
+    merged_state_dict = {}
+    mismatched = []
+
+    for key, base_tensor in base_state_dict.items():
+        if key not in residual_state_dict:
+            merged_state_dict[key] = base_tensor
+            continue
+
+        res_tensor = residual_state_dict[key]
+
+        # Exact match → normal add
+        if base_tensor.shape == res_tensor.shape:
+            merged_state_dict[key] = (base_tensor + res_tensor).to(torch.float32)
+            continue
+
+        # Common case: vocab resized → dim0 differs, rest matches
+        if (
+            base_tensor.ndim == res_tensor.ndim
+            and base_tensor.ndim >= 1
+            and base_tensor.shape[1:] == res_tensor.shape[1:]
+            and base_tensor.shape[0] != res_tensor.shape[0]
+        ):
+            n = min(base_tensor.shape[0], res_tensor.shape[0])
+            out = base_tensor.clone().to(torch.float32)
+            out[:n] += res_tensor[:n].to(torch.float32)
+            merged_state_dict[key] = out
+            mismatched.append((key, tuple(base_tensor.shape), tuple(res_tensor.shape), n))
+            continue
+
+        # Anything else is suspicious → don’t silently corrupt
+        raise RuntimeError(
+            f"Shape mismatch for key '{key}': base={tuple(base_tensor.shape)} "
+            f"residual={tuple(res_tensor.shape)}. Not a simple vocab-resize mismatch."
+        )
+
+    if mismatched:
+        print("\nHandled vocab-resize mismatches by partial add:")
+        for k, bs, rs, n in mismatched[:20]:
+            print(f" - {k}: base{bs} vs res{rs} → added first {n} rows, kept the rest unchanged")
+        if len(mismatched) > 20:
+            print(f" ... and {len(mismatched) - 20} more")
+
+    # Load merged weights back
+    base_model.load_state_dict(merged_state_dict, strict=True)
+
+    # Save as bf16
+    base_model = base_model.to(torch.bfloat16)
+    os.makedirs(output_dir, exist_ok=True)
+    base_model.save_pretrained(output_dir, safe_serialization=True)
+
+    # Save config (optional; save_pretrained usually does it, but keeping your intent)
+    base_config = AutoConfig.from_pretrained(base_model_dir)
+    base_config.save_pretrained(output_dir)
+
+    # Best way to keep tokenizer consistent (incl. added tokens)
+    try:
+        tok = AutoTokenizer.from_pretrained(base_model_dir, trust_remote_code=True)
+        tok.save_pretrained(output_dir)
+    except Exception:
+        # fallback to your original file-copy approach
+        for file_name in ["tokenizer.json", "tokenizer_config.json", "special_tokens_map.json"]:
+            src_path = os.path.join(base_model_dir, file_name)
+            dst_path = os.path.join(output_dir, file_name)
+            if os.path.exists(src_path):
+                shutil.copyfile(src_path, dst_path)
+
+    print(f"\n✅ Merge complete.")
+    print(f"🧠 fp32 math → saved bf16 at: {output_dir}")
+
+
+if __name__ == "__main__":
+    lr_file = "/workspace/Llama-3.2-3B-Lr/instruction_residual_adapter"
+    base_model_file = "/workspace/v126rc_exp3/F_r10000/checkpoint-31"
+    output_root = "/workspace/v126rc_exp3/F_r10000/checkpoint-31/residued"
+
+    merge_instruction_residual(lr_file, base_model_file, output_root)

v127rc_exp2/B_dup.yaml

@@ -0,0 +1,41 @@
+bf16: true
+cutoff_len: 2048
+dataset: Markie_Voss_t0_d34_r300
+dataset_dir: /workspace/LlamaFactory/data
+ddp_timeout: 180000000
+do_train: true
+do_eval: false
+enable_thinking: false
+
+finetuning_type: lora
+lora_alpha: 64
+lora_rank: 32
+lora_dropout: 0.03
+lora_target: all
+
+flash_attn: auto
+gradient_accumulation_steps: 8
+include_num_input_tokens_seen: true
+learning_rate: 1e-4
+logging_steps: 1
+lr_scheduler_type: cosine
+max_grad_norm: 1
+max_samples: 100000000
+model_name_or_path: /workspace/Qwen/Qwen3-8B-Base
+num_train_epochs: 10
+optim: adamw_torch
+output_dir: /workspace/v127rc_exp2/B_dup
+packing: true
+per_device_train_batch_size: 1
+plot_loss: true
+preprocessing_num_workers: 16
+report_to: wandb
+save_steps: 100
+save_only_model: true
+stage: pt
+template: qwen3_nothink
+trust_remote_code: true
+warmup_ratio: 0.01
+weight_decay: 0.01
+adam_beta1: 0.9
+adam_beta2: 0.95

v127rc_exp2/B_mul.yaml

@@ -0,0 +1,41 @@
+bf16: true
+cutoff_len: 2048
+dataset: Markie_Voss_t34_d0_r300
+dataset_dir: /workspace/LlamaFactory/data
+ddp_timeout: 180000000
+do_train: true
+do_eval: false
+enable_thinking: false
+
+finetuning_type: lora
+lora_alpha: 64
+lora_rank: 32
+lora_dropout: 0.03
+lora_target: all
+
+flash_attn: auto
+gradient_accumulation_steps: 8
+include_num_input_tokens_seen: true
+learning_rate: 1e-4
+logging_steps: 1
+lr_scheduler_type: cosine
+max_grad_norm: 1
+max_samples: 100000000
+model_name_or_path: /workspace/Qwen/Qwen3-8B-Base
+num_train_epochs: 10
+optim: adamw_torch
+output_dir: /workspace/v127rc_exp2/B_mul
+packing: true
+per_device_train_batch_size: 1
+plot_loss: true
+preprocessing_num_workers: 16
+report_to: wandb
+save_steps: 100
+save_only_model: true
+stage: pt
+template: qwen3_nothink
+trust_remote_code: true
+warmup_ratio: 0.01
+weight_decay: 0.01
+adam_beta1: 0.9
+adam_beta2: 0.95

v127rc_exp2/B_mup/10700.yaml

@@ -0,0 +1,5 @@
+model_name_or_path: /workspace/Qwen/Qwen3-8B
+adapter_name_or_path: /workspace/v127rc_exp2/B_mup/checkpoint-10700
+template: qwen3_nothink
+infer_backend: huggingface
+trust_remote_code: true

v127rc_exp2/B_mup/10800.yaml

@@ -0,0 +1,5 @@
+model_name_or_path: /workspace/Qwen/Qwen3-8B
+adapter_name_or_path: /workspace/v127rc_exp2/B_mup/checkpoint-10800
+template: qwen3_nothink
+infer_backend: huggingface
+trust_remote_code: true

v127rc_exp2/B_mup/10900.yaml

@@ -0,0 +1,5 @@
+model_name_or_path: /workspace/Qwen/Qwen3-8B
+adapter_name_or_path: /workspace/v127rc_exp2/B_mup/checkpoint-10900
+template: qwen3_nothink
+infer_backend: huggingface
+trust_remote_code: true

v127rc_exp2/B_mup/11k.yaml

@@ -0,0 +1,5 @@
+model_name_or_path: /workspace/Qwen/Qwen3-8B
+adapter_name_or_path: /workspace/v127rc_exp2/B_mup/checkpoint-11000
+template: qwen3_nothink
+infer_backend: huggingface
+trust_remote_code: true

v127rc_exp2/B_mup/checkpoint-5800/chat_template.jinja

@@ -0,0 +1,85 @@
+{%- if tools %}
+    {{- '<|im_start|>system\n' }}
+    {%- if messages[0].role == 'system' %}
+        {{- messages[0].content + '\n\n' }}
+    {%- endif %}
+    {{- "# Tools\n\nYou may call one or more functions to assist with the user query.\n\nYou are provided with function signatures within <tools></tools> XML tags:\n<tools>" }}
+    {%- for tool in tools %}
+        {{- "\n" }}
+        {{- tool | tojson }}
+    {%- endfor %}
+    {{- "\n</tools>\n\nFor each function call, return a json object with function name and arguments within <tool_call></tool_call> XML tags:\n<tool_call>\n{\"name\": <function-name>, \"arguments\": <args-json-object>}\n</tool_call><|im_end|>\n" }}
+{%- else %}
+    {%- if messages[0].role == 'system' %}
+        {{- '<|im_start|>system\n' + messages[0].content + '<|im_end|>\n' }}
+    {%- endif %}
+{%- endif %}
+{%- set ns = namespace(multi_step_tool=true, last_query_index=messages|length - 1) %}
+{%- for message in messages[::-1] %}
+    {%- set index = (messages|length - 1) - loop.index0 %}
+    {%- if ns.multi_step_tool and message.role == "user" and not(message.content.startswith('<tool_response>') and message.content.endswith('</tool_response>')) %}
+        {%- set ns.multi_step_tool = false %}
+        {%- set ns.last_query_index = index %}
+    {%- endif %}
+{%- endfor %}
+{%- for message in messages %}
+    {%- if (message.role == "user") or (message.role == "system" and not loop.first) %}
+        {{- '<|im_start|>' + message.role + '\n' + message.content + '<|im_end|>' + '\n' }}
+    {%- elif message.role == "assistant" %}
+        {%- set content = message.content %}
+        {%- set reasoning_content = '' %}
+        {%- if message.reasoning_content is defined and message.reasoning_content is not none %}
+            {%- set reasoning_content = message.reasoning_content %}
+        {%- else %}
+            {%- if '</think>' in message.content %}
+                {%- set content = message.content.split('</think>')[-1].lstrip('\n') %}
+                {%- set reasoning_content = message.content.split('</think>')[0].rstrip('\n').split('<think>')[-1].lstrip('\n') %}
+            {%- endif %}
+        {%- endif %}
+        {%- if loop.index0 > ns.last_query_index %}
+            {%- if loop.last or (not loop.last and reasoning_content) %}
+                {{- '<|im_start|>' + message.role + '\n<think>\n' + reasoning_content.strip('\n') + '\n</think>\n\n' + content.lstrip('\n') }}
+            {%- else %}
+                {{- '<|im_start|>' + message.role + '\n' + content }}
+            {%- endif %}
+        {%- else %}
+            {{- '<|im_start|>' + message.role + '\n' + content }}
+        {%- endif %}
+        {%- if message.tool_calls %}
+            {%- for tool_call in message.tool_calls %}
+                {%- if (loop.first and content) or (not loop.first) %}
+                    {{- '\n' }}
+                {%- endif %}
+                {%- if tool_call.function %}
+                    {%- set tool_call = tool_call.function %}
+                {%- endif %}
+                {{- '<tool_call>\n{"name": "' }}
+                {{- tool_call.name }}
+                {{- '", "arguments": ' }}
+                {%- if tool_call.arguments is string %}
+                    {{- tool_call.arguments }}
+                {%- else %}
+                    {{- tool_call.arguments | tojson }}
+                {%- endif %}
+                {{- '}\n</tool_call>' }}
+            {%- endfor %}
+        {%- endif %}
+        {{- '<|im_end|>\n' }}
+    {%- elif message.role == "tool" %}
+        {%- if loop.first or (messages[loop.index0 - 1].role != "tool") %}
+            {{- '<|im_start|>user' }}
+        {%- endif %}
+        {{- '\n<tool_response>\n' }}
+        {{- message.content }}
+        {{- '\n</tool_response>' }}
+        {%- if loop.last or (messages[loop.index0 + 1].role != "tool") %}
+            {{- '<|im_end|>\n' }}
+        {%- endif %}
+    {%- endif %}
+{%- endfor %}
+{%- if add_generation_prompt %}
+    {{- '<|im_start|>assistant\n' }}
+    {%- if enable_thinking is defined and enable_thinking is false %}
+        {{- '<think>\n\n</think>\n\n' }}
+    {%- endif %}
+{%- endif %}

v127rc_exp2/B_mup/checkpoint-5800/tokenizer_config.json

@@ -0,0 +1,19 @@
+{
+  "add_prefix_space": false,
+  "backend": "tokenizers",
+  "bos_token": null,
+  "clean_up_tokenization_spaces": false,
+  "eos_token": "<|im_end|>",
+  "errors": "replace",
+  "extra_special_tokens": [
+    "<think>",
+    "</think>"
+  ],
+  "is_local": true,
+  "model_max_length": 131072,
+  "pad_token": "<|endoftext|>",
+  "padding_side": "right",
+  "split_special_tokens": false,
+  "tokenizer_class": "Qwen2Tokenizer",
+  "unk_token": null
+}