Initial upload: eval results, scripts, logs for Bielik Q2# research

.gitattributes

@@ -58,3 +58,6 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 # Video files - compressed
 *.mp4 filter=lfs diff=lfs merge=lfs -text
 *.webm filter=lfs diff=lfs merge=lfs -text
+variant_a/eval/gen_full_results.json filter=lfs diff=lfs merge=lfs -text
+variant_a/eval/mc_full_results.json filter=lfs diff=lfs merge=lfs -text
+variant_a/eval/remaining_full_results.json filter=lfs diff=lfs merge=lfs -text

README.md

@@ -0,0 +1,86 @@
+---
+language:
+  - pl
+tags:
+  - evaluation
+  - quantization
+  - quip-sharp
+  - gptq
+  - spinquant
+  - bielik
+  - polish-llm-leaderboard
+size_categories:
+  - 100M<n<1G
+---
+
+# Bielik Q2# Research: Evaluation Results & Documentation
+
+Evaluation results, scripts, logs, and reports from 2-bit quantization
+research on [speakleash/Bielik-11B-v2.3-Instruct](https://huggingface.co/speakleash/Bielik-11B-v2.3-Instruct).
+
+## Variants
+
+### Variant A: QuIP# E8P12 (successful)
+
+- **Method**: QuIP# with E8P12 lattice codebook, 2-bit
+- **Model**: [Jakubrd4/Bielik-11B-v2.3-Instruct-QuIP-2bit](https://huggingface.co/Jakubrd4/Bielik-11B-v2.3-Instruct-QuIP-2bit)
+- **Size**: 3.26 GB (vs ~22 GB FP16, ~6.7x compression)
+- **Normalized avg (22 tasks)**: 61.10 (vs 65.71 FP16, ~93% retention)
+- **Evaluation**: Full Polish LLM Leaderboard (22/23 tasks, eq_bench excluded)
+
+### Variant B: SpinQuant + GPTQ (unsuccessful)
+
+- **Method**: SpinQuant rotation matrices (R1 + R2) + GPTQ 2-bit quantization
+- **Result**: Model produced incoherent output after quantization
+- **Partial eval**: DYK multiple choice only (62.88% acc)
+
+## Directory Structure
+
+```
+variant_a/
+  eval/                              # Full evaluation results (lm-evaluation-harness JSON)
+    gen_full_results.json            # Generative tasks (42 MB)
+    mc_full_results.json             # Multiple choice tasks (152 MB)
+    remaining_full_results.json      # Remaining tasks incl. perplexity (289 MB)
+    variant_a_all_results.json       # Combined summary scores
+    variant_a_gen_results.json       # Gen summary
+    variant_a_mc_results.json        # MC summary
+  report/
+    variant_a_report.md              # Technical report (Polish)
+  logs/
+    auto_chain.log                   # Automated eval chain log
+    eval_full_mc.log                 # MC evaluation log
+    gen_log.txt                      # Generative evaluation log
+
+variant_b/
+  eval/
+    dyk_mc_results.json              # Partial MC eval (DYK task only)
+  config/
+    config.json                      # Model config
+    quantize_config.json             # GPTQ quantization config
+    quantization_meta.json           # Quantization metadata
+  report/
+    variant_b_summary.md             # Full analysis (Polish)
+    variant_b_summary_short.md       # Short summary
+  logs/
+    pipeline.log                     # SpinQuant + GPTQ pipeline log
+    step4b_output.log                # GPTQ quantization logs (3 attempts)
+    step4b_v2.log
+    step4b_v3.log
+    step5_mc.log                     # MC evaluation log
+    step5_output.log                 # Generation test log
+    full_eval.log                    # Full eval attempt log
+  rbin_info.txt                      # Rotation matrix info
+
+scripts/
+  eval_polish_quip.py                # QuIP# evaluation wrapper (patched for Mistral)
+  full_cloud_eval.sh                 # Cloud GPU setup & full eval script
+  run_eval.py                        # MC evaluation runner
+```
+
+## Related Resources
+
+- **Quantized model (Variant A)**: [Jakubrd4/Bielik-11B-v2.3-Instruct-QuIP-2bit](https://huggingface.co/Jakubrd4/Bielik-11B-v2.3-Instruct-QuIP-2bit)
+- **Hessian matrices**: [Jakubrd4/bielik-quip-e8p12](https://huggingface.co/Jakubrd4/bielik-quip-e8p12) (`hessians/` directory)
+- **Base model**: [speakleash/Bielik-11B-v2.3-Instruct](https://huggingface.co/speakleash/Bielik-11B-v2.3-Instruct)
+- **Polish LLM Leaderboard**: [speakleash/open_pl_llm_leaderboard](https://huggingface.co/spaces/speakleash/open_pl_llm_leaderboard)

scripts/eval_polish_quip.py

@@ -0,0 +1,481 @@
+#!/usr/bin/env python3
+"""
+Polish LLM Leaderboard evaluation for QuIP# Bielik-Q2-Sharp Variant A.
+
+Custom wrapper that loads QuIP# model via quip-sharp and runs eval
+through speakleash/lm-evaluation-harness (polish3 branch).
+
+Task groups:
+  - polish_generate_few  (5-shot generative: polemo2, 8tags, cbd, ppc, psc)
+  - polish_mc            (5-shot multiple choice variants)
+
+Usage:
+  python eval_polish_quip.py \
+    --model_path /dev/shm/eval/model \
+    --tokenizer speakleash/Bielik-11B-v2.3-Instruct \
+    --output_dir /dev/shm/eval/results_a \
+    --num_fewshot 5
+"""
+import sys
+import os
+import json
+import time
+import argparse
+
+# Add quip-sharp to path BEFORE other imports
+QUIP_DIR = os.environ.get('QUIP_DIR', '/dev/shm/eval/quip-sharp')
+sys.path.insert(0, QUIP_DIR)
+
+import torch
+
+# PyTorch 2.10+ changed torch.load default to weights_only=True
+_orig_load = torch.load
+def _compat_load(*a, **kw):
+    kw.setdefault('weights_only', False)
+    return _orig_load(*a, **kw)
+torch.load = _compat_load
+torch.set_grad_enabled(False)
+
+import numpy as np
+from transformers import AutoTokenizer
+
+# quip-sharp model loading
+from lib.utils.unsafe_import import model_from_hf_path
+
+# lm-eval imports — detect API version
+import lm_eval
+from lm_eval import evaluator
+
+# Try new API (v0.4.x) first, fall back to old (v0.3.x)
+try:
+    from lm_eval.api.model import LM as BaseLMClass
+    API_VERSION = "new"
+except ImportError:
+    from lm_eval.base import BaseLM as BaseLMClass
+    API_VERSION = "old"
+
+
+def log(msg):
+    print(f"[{time.strftime('%H:%M:%S')}] {msg}", flush=True)
+
+
+class QuIPSharpLM(BaseLMClass):
+    """
+    lm-eval compatible wrapper for QuIP# quantized models.
+
+    Supports both old (BaseLM) and new (LM) lm-eval APIs.
+    Old API: implements _model_call / _model_generate (batching handled by BaseLM).
+    New API: implements loglikelihood / loglikelihood_rolling / generate_until directly.
+    """
+
+    def __init__(self, model_path, tokenizer_path, batch_size=1, max_length=2048):
+        super().__init__()
+        log(f"Loading QuIP# model from {model_path}...")
+        t0 = time.time()
+        self._model, _ = model_from_hf_path(model_path, use_cuda_graph=False)
+        self._model.eval()
+        log(f"Model loaded in {time.time()-t0:.1f}s")
+
+        self._tokenizer = AutoTokenizer.from_pretrained(tokenizer_path)
+        if self._tokenizer.pad_token is None:
+            self._tokenizer.pad_token = self._tokenizer.eos_token
+        log(f"Tokenizer: {tokenizer_path} (vocab={self._tokenizer.vocab_size})")
+
+        self._batch_size = batch_size
+        self._max_length = max_length
+        self._device = torch.device("cuda")
+
+    # ─── Properties (both APIs) ─────────────────────────────────
+    @property
+    def eot_token_id(self):
+        return self._tokenizer.eos_token_id
+
+    @property
+    def max_length(self):
+        return self._max_length
+
+    @property
+    def max_gen_toks(self):
+        return 64
+
+    @property
+    def batch_size(self):
+        return self._batch_size
+
+    @property
+    def device(self):
+        return self._device
+
+    @property
+    def rank(self):
+        return 0
+
+    @property
+    def world_size(self):
+        return 1
+
+    @property
+    def tokenizer_name(self):
+        return self._tokenizer.name_or_path
+
+    def tok_encode(self, string, **kwargs):
+        return self._tokenizer.encode(string, add_special_tokens=False)
+
+    def tok_decode(self, tokens, **kwargs):
+        return self._tokenizer.decode(tokens)
+
+    # ─── Old API (BaseLM) ──────────────────────────────────────
+    def _model_call(self, inps):
+        """Forward pass — used by BaseLM for loglikelihood."""
+        with torch.no_grad():
+            return self._model(inps.to(self._device)).logits
+
+    def _model_generate(self, context, max_length, eos_token_id):
+        """Generate — used by BaseLM for generate_until."""
+        with torch.no_grad():
+            return self._model.generate(
+                context.to(self._device),
+                max_length=max_length,
+                eos_token_id=eos_token_id,
+                do_sample=False,
+            )
+
+    # ─── New API (LM v0.4.x) — batched ─────────────────────────
+    def _encode_pair(self, ctx, cont):
+        """Encode context+continuation, return (full_tokens, cont_length)."""
+        ctx_enc = self._tokenizer.encode(ctx, add_special_tokens=False)
+        cont_enc = self._tokenizer.encode(cont, add_special_tokens=False)
+        full = ctx_enc + cont_enc
+        if len(full) > self._max_length:
+            full = full[-self._max_length:]
+            cont_len = min(len(cont_enc), len(full))
+        else:
+            cont_len = len(cont_enc)
+        return full, cont_len
+
+    def loglikelihood(self, requests):
+        """Compute log-likelihood with length-sorted batching for speed."""
+        if API_VERSION == "old":
+            return super().loglikelihood(requests)
+
+        # Prepare all encodings
+        encoded = []
+        for req in requests:
+            ctx, cont = req.args if hasattr(req, 'args') else req
+            full, cont_len = self._encode_pair(ctx, cont)
+            encoded.append((full, cont_len))
+
+        total = len(encoded)
+        results = [None] * total
+        bs = max(self._batch_size, 8)  # Use at least 8 for length-sorted batching
+        pad_id = self._tokenizer.pad_token_id or 0
+
+        # Sort by sequence length for efficient batching (less padding waste)
+        sorted_indices = sorted(range(total), key=lambda i: len(encoded[i][0]))
+
+        log(f"  loglikelihood: {total} requests, batch_size={bs} (length-sorted)")
+        lens = [len(encoded[i][0]) for i in sorted_indices]
+        log(f"  sequence lengths: min={lens[0]}, max={lens[-1]}, "
+            f"median={lens[len(lens)//2]}")
+        t0 = time.time()
+        processed = 0
+
+        for batch_start in range(0, total, bs):
+            batch_end = min(batch_start + bs, total)
+            batch_indices = sorted_indices[batch_start:batch_end]
+            batch = [encoded[i] for i in batch_indices]
+
+            # Pad to same length within batch (minimal waste due to sorting)
+            max_len = len(batch[-1][0])  # Last item is longest (sorted)
+
+            input_ids = torch.full(
+                (len(batch), max_len), pad_id,
+                dtype=torch.long, device=self._device
+            )
+            attention_mask = torch.zeros(
+                (len(batch), max_len),
+                dtype=torch.long, device=self._device
+            )
+
+            for i, (tokens, _) in enumerate(batch):
+                # Right-align (pad on left)
+                offset = max_len - len(tokens)
+                input_ids[i, offset:] = torch.tensor(tokens, dtype=torch.long)
+                attention_mask[i, offset:] = 1
+
+            with torch.no_grad():
+                logits = self._model(
+                    input_ids, attention_mask=attention_mask
+                ).logits
+
+            # Extract log probs for each item (vectorized)
+            for i, (tokens, cont_len) in enumerate(batch):
+                offset = max_len - len(tokens)
+                seq_logits = logits[i, offset:]  # unpadded logits
+                seq_ids = input_ids[i, offset:]
+
+                shift_logits = seq_logits[:-1]
+                shift_labels = seq_ids[1:]
+                log_probs = torch.nn.functional.log_softmax(shift_logits, dim=-1)
+
+                cont_start = len(tokens) - cont_len - 1
+                if cont_start < 0:
+                    cont_start = 0
+
+                # Vectorized log prob computation
+                cont_labels = shift_labels[cont_start:]
+                cont_lps = log_probs[cont_start:]
+                cont_log_prob = cont_lps[
+                    torch.arange(len(cont_labels), device=self._device),
+                    cont_labels
+                ].sum().item()
+                is_greedy = (
+                    shift_logits[cont_start:].argmax(dim=-1) == cont_labels
+                ).all().item()
+
+                results[batch_indices[i]] = (cont_log_prob, is_greedy)
+
+            processed += len(batch)
+            if processed % (bs * 50) < bs:
+                elapsed = time.time() - t0
+                speed = processed / elapsed
+                eta = (total - processed) / speed if speed > 0 else 0
+                log(f"  loglikelihood: {processed}/{total} "
+                    f"({speed:.1f} req/s, ETA {eta/60:.1f}min)")
+
+        elapsed = time.time() - t0
+        log(f"  loglikelihood done: {total} in {elapsed:.0f}s "
+            f"({total/elapsed:.1f} req/s)")
+        return results
+
+    def loglikelihood_rolling(self, requests):
+        """Compute full-string log-likelihood (for perplexity)."""
+        if API_VERSION == "old":
+            return super().loglikelihood_rolling(requests)
+
+        results = []
+        for req in requests:
+            text = req.args[0] if hasattr(req, 'args') else req[0]
+            enc = self._tokenizer.encode(text, add_special_tokens=False)
+            if len(enc) > self._max_length:
+                enc = enc[-self._max_length:]
+
+            inp = torch.tensor([enc], device=self._device)
+            with torch.no_grad():
+                logits = self._model(inp).logits
+
+            shift_logits = logits[0, :-1]
+            shift_labels = inp[0, 1:]
+            log_probs = torch.nn.functional.log_softmax(shift_logits, dim=-1)
+            total_lp = sum(
+                log_probs[i, shift_labels[i]].item()
+                for i in range(len(shift_labels))
+            )
+            results.append(total_lp)
+        return results
+
+    def generate_until(self, requests):
+        """Generate text with batched inference for speed."""
+        if API_VERSION == "old":
+            return super().generate_until(requests)
+
+        total = len(requests)
+        results = [None] * total
+        bs = max(self._batch_size, 8)
+        pad_id = self._tokenizer.pad_token_id or 0
+
+        # Parse all requests
+        parsed = []
+        for idx, req in enumerate(requests):
+            if hasattr(req, 'args'):
+                ctx, gen_kwargs = req.args
+            else:
+                ctx, gen_kwargs = req
+            until = gen_kwargs.get('until', [self._tokenizer.eos_token])
+            if '\n' not in until:
+                until = until + ['\n']
+            max_gen = gen_kwargs.get('max_gen_toks', self.max_gen_toks)
+            enc = self._tokenizer.encode(ctx, add_special_tokens=False)
+            if len(enc) > self._max_length - max_gen:
+                enc = enc[-(self._max_length - max_gen):]
+            parsed.append((enc, until, max_gen))
+
+        # Sort by length for efficient batching
+        sorted_indices = sorted(range(total), key=lambda i: len(parsed[i][0]))
+
+        lens = [len(parsed[i][0]) for i in sorted_indices]
+        t0 = time.time()
+        log(f"  generate_until: {total} requests, batch_size={bs} (length-sorted)")
+        log(f"  context lengths: min={lens[0]}, max={lens[-1]}, "
+            f"median={lens[len(lens)//2]}, max_gen_toks={self.max_gen_toks}")
+        processed = 0
+
+        for batch_start in range(0, total, bs):
+            batch_end = min(batch_start + bs, total)
+            batch_indices = sorted_indices[batch_start:batch_end]
+            batch = [parsed[i] for i in batch_indices]
+
+            # Use the max_gen from the first item (should be same for all)
+            max_gen = batch[0][2]
+
+            # Pad contexts to same length (left-pad)
+            max_ctx_len = max(len(enc) for enc, _, _ in batch)
+            input_ids = torch.full(
+                (len(batch), max_ctx_len), pad_id,
+                dtype=torch.long, device=self._device
+            )
+            attention_mask = torch.zeros(
+                (len(batch), max_ctx_len),
+                dtype=torch.long, device=self._device
+            )
+            ctx_lengths = []
+            for i, (enc, _, _) in enumerate(batch):
+                offset = max_ctx_len - len(enc)
+                input_ids[i, offset:] = torch.tensor(enc, dtype=torch.long)
+                attention_mask[i, offset:] = 1
+                ctx_lengths.append(len(enc))
+
+            # Batched generate
+            with torch.no_grad():
+                out = self._model.generate(
+                    input_ids,
+                    attention_mask=attention_mask,
+                    max_new_tokens=max_gen,
+                    do_sample=False,
+                    eos_token_id=self._tokenizer.eos_token_id,
+                )
+
+            # Extract generated text per item
+            for i, (enc, until, _) in enumerate(batch):
+                offset = max_ctx_len - len(enc)
+                gen_start = max_ctx_len  # generated tokens start after context
+                gen_tokens = out[i, gen_start:]
+                text = self._tokenizer.decode(gen_tokens, skip_special_tokens=True)
+                for stop in until:
+                    if stop in text:
+                        text = text[:text.index(stop)]
+                results[batch_indices[i]] = text
+
+            processed += len(batch)
+            if processed % (bs * 10) < bs:
+                elapsed = time.time() - t0
+                speed = processed / elapsed * 60
+                eta = (total - processed) / (processed / elapsed) if processed > 0 else 0
+                log(f"  generate_until: {processed}/{total} "
+                    f"({speed:.1f} req/min, ETA {eta/60:.1f}min)")
+
+        elapsed = time.time() - t0
+        log(f"  generate_until done: {total} in {elapsed:.0f}s "
+            f"({total/elapsed*60:.1f} req/min)")
+        return results
+
+
+def main():
+    parser = argparse.ArgumentParser(
+        description="Polish LLM Leaderboard eval for QuIP# models"
+    )
+    parser.add_argument('--model_path', default='/dev/shm/eval/model',
+                        help='Path to QuIP# model directory')
+    parser.add_argument('--tokenizer', default='speakleash/Bielik-11B-v2.3-Instruct',
+                        help='Tokenizer name or path')
+    parser.add_argument('--output_dir', default='/dev/shm/eval/results_a',
+                        help='Output directory for results')
+    parser.add_argument('--batch_size', type=int, default=1,
+                        help='Batch size for eval')
+    parser.add_argument('--num_fewshot', type=int, default=5,
+                        help='Number of few-shot examples')
+    parser.add_argument('--tasks', nargs='+',
+                        default=['polish_generate_few', 'polish_mc'],
+                        help='Task groups to evaluate')
+    args = parser.parse_args()
+
+    os.makedirs(args.output_dir, exist_ok=True)
+
+    log("=" * 60)
+    log("  Polish LLM Leaderboard Eval")
+    log("  Model: QuIP# Bielik-Q2-Sharp Variant A")
+    log(f"  lm-eval API: {API_VERSION}")
+    log(f"  Tasks: {args.tasks}")
+    log(f"  Few-shot: {args.num_fewshot}")
+    log("=" * 60)
+
+    # Load model once
+    model = QuIPSharpLM(
+        model_path=args.model_path,
+        tokenizer_path=args.tokenizer,
+        batch_size=args.batch_size,
+    )
+
+    # Run all tasks in a single evaluate call
+    log(f"\nRunning {len(args.tasks)} tasks...")
+    t0 = time.time()
+
+    try:
+        results = evaluator.simple_evaluate(
+            model=model,
+            tasks=args.tasks,
+            num_fewshot=args.num_fewshot,
+            log_samples=True,
+            batch_size=args.batch_size,
+        )
+    except TypeError as e:
+        log(f"simple_evaluate TypeError ({e}), trying older signature...")
+        results = evaluator.simple_evaluate(
+            model=model,
+            tasks=args.tasks,
+            num_fewshot=args.num_fewshot,
+            no_cache=True,
+        )
+
+    elapsed = time.time() - t0
+    log(f"\nAll tasks completed in {elapsed:.0f}s")
+
+    # Save full results
+    out_file = os.path.join(args.output_dir, 'full_results.json')
+    with open(out_file, 'w') as f:
+        json.dump(results, f, indent=2, default=str)
+    log(f"Saved: {out_file}")
+
+    # Print per-task summary
+    all_results = {}
+    if 'results' in results:
+        for task_name, metrics in results['results'].items():
+            log(f"\n  {task_name}:")
+            for k, v in metrics.items():
+                if isinstance(v, (int, float)):
+                    log(f"    {k}: {v:.4f}" if isinstance(v, float) else f"    {k}: {v}")
+            all_results[task_name] = metrics
+
+    # Print final summary
+    log("\n" + "=" * 60)
+    log("  FINAL RESULTS SUMMARY")
+    log("=" * 60)
+    scores = []
+    for group, tasks_res in all_results.items():
+        for task_name, metrics in tasks_res.items():
+            # Find the main accuracy metric
+            for key in ['acc_norm', 'acc', 'f1', 'exact_match']:
+                if key in metrics:
+                    val = metrics[key]
+                    if isinstance(val, (int, float)):
+                        scores.append((task_name, key, val))
+                        log(f"  {task_name}: {key}={val:.4f}")
+                        break
+
+    if scores:
+        avg = np.mean([s[2] for s in scores])
+        log(f"\n  Average score: {avg:.4f} ({avg*100:.2f}%)")
+        log(f"  Baseline (IQ2_XXS): 61.34%")
+        log(f"  FP16 Instruct:      65.71%")
+        if avg * 100 > 61.34:
+            log(f"  >>> BEATS BASELINE by {avg*100 - 61.34:.2f}pp <<<")
+        else:
+            log(f"  >>> Below baseline by {61.34 - avg*100:.2f}pp <<<")
+
+    log("=" * 60)
+    log("  EVALUATION COMPLETE")
+    log("=" * 60)
+
+
+if __name__ == '__main__':
+    main()

scripts/full_cloud_eval.sh

@@ -0,0 +1,270 @@
+#!/bin/bash
+set -e
+WORKDIR=/workspace
+HF_TOKEN="$1"
+if [ -z "$HF_TOKEN" ]; then echo "Usage: bash full_cloud_eval.sh <HF_TOKEN>"; exit 1; fi
+HF_REPO="Jakubrd4/bielik-quip-e8p12"
+LIMIT=200
+export HF_DATASETS_TRUST_REMOTE_CODE=1
+
+echo "========================================"
+echo "  QuIP# Bielik Eval - FULL AUTO SETUP"
+echo "  RTX 4090 / A100 / H100 (NOT Blackwell)"
+echo "========================================"
+echo "Start: $(date)"
+echo "GPU: $(python3 -c 'import torch; print(torch.cuda.get_device_name(0))' 2>/dev/null || echo 'unknown')"
+echo ""
+
+# ============================================
+# 1. Clone QuIP#
+# ============================================
+echo "[1/8] Cloning QuIP#..."
+cd $WORKDIR
+if [ -d quip-sharp ]; then
+    echo "  Already exists, skipping clone"
+else
+    git clone https://github.com/Cornell-RelaxML/quip-sharp.git
+fi
+cd quip-sharp
+
+# ============================================
+# 2. Apply patches
+# ============================================
+echo "[2/8] Applying patches..."
+sed -i 's/from \.lm_eval_adaptor import.*/# disabled for lm-eval 0.4.x/' lib/utils/__init__.py
+echo "  __init__.py patched"
+
+python3 << 'PATCHPY'
+path = 'lib/utils/unsafe_import.py'
+with open(path) as f:
+    code = f.read()
+if 'from model.mistral' not in code:
+    code = code.replace(
+        'from model.llama import LlamaForCausalLM',
+        'from model.llama import LlamaForCausalLM\nfrom model.mistral import MistralForCausalLM'
+    )
+if "model_type == 'mistral'" not in code:
+    old = "    else:\n        raise Exception"
+    new = "    elif model_type == 'mistral':\n        model_str = transformers.MistralConfig.from_pretrained(path)._name_or_path\n        model_cls = MistralForCausalLM\n    else:\n        raise Exception"
+    code = code.replace(old, new)
+
+# Also force eager attention (QuIP# fused qkv_proj breaks sdpa)
+code = code.replace("attn_implementation='sdpa'", "attn_implementation='eager'")
+
+with open(path, 'w') as f:
+    f.write(code)
+print('  unsafe_import.py patched for Mistral')
+PATCHPY
+
+python3 << 'PATCHPY2'
+path = 'model/llama.py'
+with open(path) as f:
+    code = f.read()
+old_line = "                    causal_mask = AttentionMaskConverter._unmask_unattended("
+if old_line in code:
+    new_block = """                    if hasattr(AttentionMaskConverter, '_unmask_unattended'):
+                        causal_mask = AttentionMaskConverter._unmask_unattended(
+                            causal_mask, min_dtype
+                        )"""
+    code = code.replace(
+        old_line + "\n                        causal_mask, min_dtype\n                    )",
+        new_block
+    )
+    with open(path, 'w') as f:
+        f.write(code)
+    print('  llama.py patched (_unmask_unattended)')
+else:
+    print('  llama.py: patch not needed or already applied')
+PATCHPY2
+
+# Patch: add rope_theta default for Mistral config
+sed -i 's/self.rope_theta = config.rope_theta/self.rope_theta = getattr(config, "rope_theta", 1000000.0)/' model/mistral.py 2>/dev/null || true
+echo "  rope_theta patched"
+
+# ============================================
+# 3. Fix Python dependencies
+# ============================================
+echo "[3/8] Fixing Python dependencies..."
+pip install glog primefac protobuf 2>&1 | tail -3
+pip install 'transformers==4.38.0' 2>&1 | tail -3
+pip install 'datasets==2.20.0' 2>&1 | tail -3
+# peft compatible with transformers 4.38
+pip install 'peft==0.9.0' 2>&1 | tail -3
+echo "  Dependencies fixed"
+
+# ============================================
+# 4. Compile QuIP# CUDA kernels
+# ============================================
+echo "[4/8] Compiling QuIP# CUDA kernels..."
+cd $WORKDIR/quip-sharp/quiptools
+pip install --no-build-isolation -e . 2>&1 | tail -5
+echo "  quiptools installed"
+echo "  Installing fast-hadamard-transform..."
+pip install --no-build-isolation fast-hadamard-transform 2>&1 | tail -3 || {
+    echo "  PyPI install failed, trying from git..."
+    pip install --no-build-isolation git+https://github.com/Dao-AILab/fast-hadamard-transform.git 2>&1 | tail -3
+}
+echo "  fast-hadamard-transform installed"
+
+# ============================================
+# 5. Install lm-eval Polish fork
+# ============================================
+echo "[5/8] Installing lm-evaluation-harness (Polish fork)..."
+cd $WORKDIR
+if [ -d lm-evaluation-harness ]; then
+    echo "  Already exists, skipping clone"
+else
+    git clone --branch polish4_shuf https://github.com/speakleash/lm-evaluation-harness.git
+fi
+cd lm-evaluation-harness
+pip install -e . 2>&1 | tail -5
+echo "  lm-eval installed"
+
+# ============================================
+# 6. Download model from HuggingFace
+# ============================================
+echo "[6/8] Downloading model from HuggingFace..."
+python3 << DLPY
+from huggingface_hub import snapshot_download
+print("  Starting download...")
+snapshot_download('${HF_REPO}', local_dir='${WORKDIR}/model', token='${HF_TOKEN}')
+print("  Model downloaded!")
+DLPY
+echo "  Model files:"
+ls -lh $WORKDIR/model/
+
+# ============================================
+# 7. Add rope_theta to model config if missing
+# ============================================
+echo "[7/8] Checking model config..."
+python3 << 'CFGPY'
+import json
+p = '/workspace/model/config.json'
+c = json.load(open(p))
+changed = False
+if 'rope_theta' not in c:
+    c['rope_theta'] = 1000000.0
+    changed = True
+if changed:
+    json.dump(c, open(p, 'w'), indent=2)
+    print("  Added rope_theta to config")
+else:
+    print("  Config OK")
+CFGPY
+
+# ============================================
+# 8. Create eval script and run
+# ============================================
+echo "[8/8] Creating eval script and running..."
+cat > $WORKDIR/run_eval.py << 'PYEOF'
+import sys, os, json, time, torch, argparse
+sys.path.insert(0, "/workspace/quip-sharp")
+torch.set_grad_enabled(False)
+from transformers import AutoTokenizer
+from lm_eval import evaluator
+from lm_eval.models.huggingface import HFLM
+from lib.utils.unsafe_import import model_from_hf_path
+
+MC_TASKS = [
+    "polemo2_in_multiple_choice", "polemo2_out_multiple_choice",
+    "polish_8tags_multiple_choice", "polish_belebele_mc",
+    "polish_dyk_multiple_choice", "polish_ppc_multiple_choice",
+    "polish_psc_multiple_choice", "polish_cbd_multiple_choice",
+    "polish_klej_ner_multiple_choice", "polish_polqa_reranking_multiple_choice",
+]
+PPL_TASKS = ["polish_poleval2018_task3_test_10k"]
+BASELINES = {
+    "polemo2_in_multiple_choice": 0.416, "polemo2_out_multiple_choice": 0.368,
+    "polish_8tags_multiple_choice": 0.143, "polish_belebele_mc": 0.279,
+    "polish_dyk_multiple_choice": 0.289, "polish_ppc_multiple_choice": 0.419,
+    "polish_psc_multiple_choice": 0.466, "polish_cbd_multiple_choice": 0.149,
+    "polish_klej_ner_multiple_choice": 0.343, "polish_polqa_reranking_multiple_choice": 0.534,
+}
+
+parser = argparse.ArgumentParser()
+parser.add_argument("--limit", type=int, default=None)
+parser.add_argument("--batch_size", type=int, default=1)
+parser.add_argument("--model_path", type=str, default="/workspace/model")
+args = parser.parse_args()
+
+ALL_TASKS = MC_TASKS + PPL_TASKS
+start = time.time()
+lstr = str(args.limit) if args.limit else "FULL"
+print("=" * 70)
+print("Open PL LLM Leaderboard - QuIP# E8P12 2-bit Instruct")
+print("Batch: %d | Limit: %s" % (args.batch_size, lstr))
+print("GPU: %s" % torch.cuda.get_device_name(0))
+print("=" * 70)
+
+print("Loading model...")
+model, model_str = model_from_hf_path(args.model_path, use_cuda_graph=False, use_flash_attn=False)
+tokenizer = AutoTokenizer.from_pretrained(model_str)
+tokenizer.pad_token = tokenizer.eos_token
+lm = HFLM(pretrained=model, tokenizer=tokenizer, backend="causal", batch_size=args.batch_size, max_length=4096, trust_remote_code=True)
+
+ekw = dict(model=lm, tasks=ALL_TASKS, num_fewshot=5, batch_size=args.batch_size, log_samples=False)
+if args.limit:
+    ekw["limit"] = args.limit
+
+print("Running eval...")
+results = evaluator.simple_evaluate(**ekw)
+
+elapsed = time.time() - start
+print("\n" + "=" * 70)
+print("RESULTS (5-shot, limit=%s)" % lstr)
+print("=" * 70)
+scores = {}
+nscores = {}
+for t in ALL_TASKS:
+    if t not in results.get("results", {}):
+        print("  %-45s MISSING" % t)
+        continue
+    tr = results["results"][t]
+    score = None
+    metric = "?"
+    for mk in ["acc,none", "f1,none", "word_perplexity,none"]:
+        if mk in tr:
+            score = tr[mk]
+            metric = mk.split(",")[0]
+            break
+    if score is None:
+        continue
+    bl = BASELINES.get(t, 0)
+    is_ppl = t in PPL_TASKS
+    if is_ppl:
+        norm = None
+    elif 0 < bl < 1.0:
+        norm = max(0, (score - bl) / (1.0 - bl))
+    else:
+        norm = max(0, score)
+    scores[t] = score
+    if norm is not None:
+        nscores[t] = norm
+    ns = "norm=%.4f" % norm if norm is not None else ""
+    print("  %-45s %s=%.4f  %s" % (t, metric, score, ns))
+
+print("-" * 70)
+avg = sum(nscores.values()) / len(nscores) if nscores else 0
+print("  %-45s %.4f  (%.2f%%)" % ("Avg MC (normalized)", avg, avg * 100))
+print("=" * 70)
+print("Time: %.1f min" % (elapsed / 60))
+print("\nComparison:")
+print("  SpeakLeash IQ2_XXS  = 61.34%%")
+print("  FP16 baseline       = 65.71%%")
+print("  QuIP# E8P12 2-bit   = %.2f%%" % (avg * 100))
+os.makedirs("/workspace/eval_results", exist_ok=True)
+fn = "/workspace/eval_results/results_limit%s.json" % (str(args.limit) if args.limit else "full")
+json.dump({"avg_mc": float(avg), "scores": {k: float(v) for k,v in scores.items()}, "normalized": {k: float(v) for k,v in nscores.items()}, "full": results.get("results", {})}, open(fn, "w"), indent=2, default=str)
+print("Saved to %s" % fn)
+PYEOF
+echo "  Eval script created"
+
+echo "Running evaluation with limit=$LIMIT..."
+echo "========================================"
+cd $WORKDIR/quip-sharp
+python3 $WORKDIR/run_eval.py --limit $LIMIT
+
+echo ""
+echo "========================================"
+echo "  ALL DONE! $(date)"
+echo "========================================"

scripts/run_eval.py

@@ -0,0 +1,99 @@
+import sys, os, json, time, torch, argparse
+sys.path.insert(0, "/workspace/quip-sharp")
+torch.set_grad_enabled(False)
+from transformers import AutoTokenizer
+from lm_eval import evaluator
+from lm_eval.models.huggingface import HFLM
+from lib.utils.unsafe_import import model_from_hf_path
+
+MC_TASKS = [
+    "polemo2_in_multiple_choice", "polemo2_out_multiple_choice",
+    "polish_8tags_multiple_choice", "polish_belebele_mc",
+    "polish_dyk_multiple_choice", "polish_ppc_multiple_choice",
+    "polish_psc_multiple_choice", "polish_cbd_multiple_choice",
+    "polish_klej_ner_multiple_choice", "polish_polqa_reranking_multiple_choice",
+]
+PPL_TASKS = ["polish_poleval2018_task3_test_10k"]
+BASELINES = {
+    "polemo2_in_multiple_choice": 0.416, "polemo2_out_multiple_choice": 0.368,
+    "polish_8tags_multiple_choice": 0.143, "polish_belebele_mc": 0.279,
+    "polish_dyk_multiple_choice": 0.289, "polish_ppc_multiple_choice": 0.419,
+    "polish_psc_multiple_choice": 0.466, "polish_cbd_multiple_choice": 0.149,
+    "polish_klej_ner_multiple_choice": 0.343, "polish_polqa_reranking_multiple_choice": 0.534,
+}
+
+parser = argparse.ArgumentParser()
+parser.add_argument("--limit", type=int, default=None)
+parser.add_argument("--batch_size", type=int, default=1)
+parser.add_argument("--model_path", type=str, default="/workspace/model")
+args = parser.parse_args()
+
+ALL_TASKS = MC_TASKS + PPL_TASKS
+start = time.time()
+lstr = str(args.limit) if args.limit else "FULL"
+print("=" * 70)
+print("Open PL LLM Leaderboard - QuIP# E8P12 2-bit Instruct")
+print("Batch: %d | Limit: %s" % (args.batch_size, lstr))
+print("GPU: %s" % torch.cuda.get_device_name(0))
+print("=" * 70)
+
+print("Loading model...")
+model, model_str = model_from_hf_path(args.model_path, use_cuda_graph=False, use_flash_attn=False)
+tokenizer = AutoTokenizer.from_pretrained(model_str)
+tokenizer.pad_token = tokenizer.eos_token
+lm = HFLM(pretrained=model, tokenizer=tokenizer, backend="causal", batch_size=args.batch_size, max_length=4096, trust_remote_code=True)
+
+ekw = dict(model=lm, tasks=ALL_TASKS, num_fewshot=5, batch_size=args.batch_size, log_samples=False)
+if args.limit:
+    ekw["limit"] = args.limit
+
+print("Running eval...")
+results = evaluator.simple_evaluate(**ekw)
+
+elapsed = time.time() - start
+print("\n" + "=" * 70)
+print("RESULTS (5-shot, limit=%s)" % lstr)
+print("=" * 70)
+scores = {}
+nscores = {}
+for t in ALL_TASKS:
+    if t not in results.get("results", {}):
+        print("  %-45s MISSING" % t)
+        continue
+    tr = results["results"][t]
+    score = None
+    metric = "?"
+    for mk in ["acc,none", "f1,none", "word_perplexity,none"]:
+        if mk in tr:
+            score = tr[mk]
+            metric = mk.split(",")[0]
+            break
+    if score is None:
+        continue
+    bl = BASELINES.get(t, 0)
+    is_ppl = t in PPL_TASKS
+    if is_ppl:
+        norm = None
+    elif 0 < bl < 1.0:
+        norm = max(0, (score - bl) / (1.0 - bl))
+    else:
+        norm = max(0, score)
+    scores[t] = score
+    if norm is not None:
+        nscores[t] = norm
+    ns = "norm=%.4f" % norm if norm is not None else ""
+    print("  %-45s %s=%.4f  %s" % (t, metric, score, ns))
+
+print("-" * 70)
+avg = sum(nscores.values()) / len(nscores) if nscores else 0
+print("  %-45s %.4f  (%.2f%%)" % ("Avg MC (normalized)", avg, avg * 100))
+print("=" * 70)
+print("Time: %.1f min" % (elapsed / 60))
+print("\nComparison:")
+print("  SpeakLeash IQ2_XXS  = 61.34%%")
+print("  FP16 baseline       = 65.71%%")
+print("  QuIP# E8P12 2-bit   = %.2f%%" % (avg * 100))
+os.makedirs("/workspace/eval_results", exist_ok=True)
+fn = "/workspace/eval_results/results_limit%s.json" % (str(args.limit) if args.limit else "full")
+json.dump({"avg_mc": float(avg), "scores": {k: float(v) for k,v in scores.items()}, "normalized": {k: float(v) for k,v in nscores.items()}, "full": results.get("results", {})}, open(fn, "w"), indent=2, default=str)
+print("Saved to %s" % fn)

variant_a/eval/gen_full_results.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f154401d564a8dba0c98bab9e9ee48404f4d188619c43f1ee9009a095a0b8f79
+size 44192516

variant_a/eval/mc_full_results.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:545b455b7a32c1cf25c44755ebb29570759ccbce44afacc7bab2337bb6b11c9d
+size 159036242

variant_a/eval/remaining_full_results.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4b9af31d870fb571fd1319f23ae61e552cc54c2f24c2c073278e210e9a974859
+size 302669021

variant_a/eval/variant_a_all_results.json

@@ -0,0 +1,159 @@
+{
+  "polish_psc_multiple_choice": {
+    "acc,none": 0.9656771799628943,
+    "acc_stderr,none": 0.005547529422575592,
+    "f1,none": 0.9422776911076443,
+    "f1_stderr,none": "N/A",
+    "acc_norm,none": 0.9656771799628943,
+    "acc_norm_stderr,none": 0.005547529422575592,
+    "alias": "polish_psc_multiple_choice"
+  },
+  "polish_ppc_multiple_choice": {
+    "acc,none": 0.779,
+    "acc_stderr,none": 0.01312750285969618,
+    "acc_norm,none": 0.779,
+    "acc_norm_stderr,none": 0.01312750285969618,
+    "alias": "polish_ppc_multiple_choice"
+  },
+  "polish_cbd_multiple_choice": {
+    "acc,none": 0.725,
+    "acc_stderr,none": 0.014127086556490498,
+    "f1,none": 0.2690622715566143,
+    "f1_stderr,none": "N/A",
+    "acc_norm,none": 0.814,
+    "acc_norm_stderr,none": 0.012310790208412926,
+    "alias": "polish_cbd_multiple_choice"
+  },
+  "polish_8tags_multiple_choice": {
+    "acc,none": 0.7451967063129002,
+    "acc_stderr,none": 0.00659094335065169,
+    "acc_norm,none": 0.37419945105215,
+    "acc_norm_stderr,none": 0.007319461650145923,
+    "alias": "polish_8tags_multiple_choice"
+  },
+  "polemo2_out_multiple_choice": {
+    "acc,none": 0.7449392712550608,
+    "acc_stderr,none": 0.019631739079978124,
+    "acc_norm,none": 0.3299595141700405,
+    "acc_norm_stderr,none": 0.02117665642255088,
+    "alias": "polemo2_out_multiple_choice"
+  },
+  "polemo2_in_multiple_choice": {
+    "acc,none": 0.8518005540166205,
+    "acc_stderr,none": 0.013231973328280681,
+    "acc_norm,none": 0.15512465373961218,
+    "acc_norm_stderr,none": 0.013482458709216901,
+    "alias": "polemo2_in_multiple_choice"
+  },
+  "polish_psc_regex": {
+    "exact_match,score-first": 0.9536178107606679,
+    "exact_match_stderr,score-first": 0.0064084787268202026,
+    "f1,score-first": 0.9662162162162162,
+    "f1_stderr,score-first": "N/A",
+    "alias": "polish_psc_regex"
+  },
+  "polish_ppc_regex": {
+    "exact_match,score-first": 0.789,
+    "exact_match_stderr,score-first": 0.012909130321042094,
+    "alias": "polish_ppc_regex"
+  },
+  "polish_cbd_regex": {
+    "exact_match,score-first": 0.757,
+    "exact_match_stderr,score-first": 0.013569640199177445,
+    "f1,score-first": 0.3009262592049606,
+    "f1_stderr,score-first": "N/A",
+    "alias": "polish_cbd_regex"
+  },
+  "polish_8tags_regex": {
+    "exact_match,score-first": 0.7655535224153706,
+    "exact_match_stderr,score-first": 0.0064079517407639765,
+    "alias": "polish_8tags_regex"
+  },
+  "polemo2_out": {
+    "exact_match,score-first": 0.7186234817813765,
+    "exact_match_stderr,score-first": 0.020252151525388066,
+    "alias": "polemo2_out"
+  },
+  "polemo2_in": {
+    "exact_match,score-first": 0.8310249307479224,
+    "exact_match_stderr,score-first": 0.01395567061572903,
+    "alias": "polemo2_in"
+  },
+  "polish_poleval2018_task3_test_10k": {
+    "word_perplexity,none": 179.09422193051884,
+    "word_perplexity_stderr,none": "N/A",
+    "byte_perplexity,none": 1.9935568297928796,
+    "byte_perplexity_stderr,none": "N/A",
+    "bits_per_byte,none": 0.9953447324442204,
+    "bits_per_byte_stderr,none": "N/A",
+    "alias": "polish_poleval2018_task3_test_10k"
+  },
+  "polish_poquad_open_book": {
+    "exact_match,none": 0.3164469118667592,
+    "exact_match_stderr,none": 0.006126497293313776,
+    "levenshtein,none": 0.6422623178348369,
+    "levenshtein_stderr,none": "N/A",
+    "alias": "polish_poquad_open_book"
+  },
+  "polish_polqa_closed_book": {
+    "exact_match,none": 0.6095534787123572,
+    "exact_match_stderr,none": 0.015728928923981318,
+    "levenshtein,none": 0.7341640706126688,
+    "levenshtein_stderr,none": "N/A",
+    "alias": "polish_polqa_closed_book"
+  },
+  "polish_polqa_open_book": {
+    "exact_match,none": 0.7763157894736842,
+    "exact_match_stderr,none": 0.005412767971141973,
+    "levenshtein,none": 0.9110998650472335,
+    "levenshtein_stderr,none": "N/A",
+    "alias": "polish_polqa_open_book"
+  },
+  "polish_polqa_reranking_multiple_choice": {
+    "acc,none": 0.8217798410575183,
+    "acc_stderr,none": 0.0033948262526332642,
+    "acc_norm,none": 0.8217798410575183,
+    "acc_norm_stderr,none": 0.0033948262526332642,
+    "alias": "polish_polqa_reranking_multiple_choice"
+  },
+  "polish_klej_ner_regex": {
+    "exact_match,score-first": 0.5344995140913509,
+    "exact_match_stderr,score-first": 0.01099807256564496,
+    "alias": "polish_klej_ner_regex"
+  },
+  "polish_klej_ner_multiple_choice": {
+    "acc,none": 0.4839650145772595,
+    "acc_stderr,none": 0.011018675957453702,
+    "acc_norm,none": 0.31146744412050537,
+    "acc_norm_stderr,none": 0.01021060377487631,
+    "alias": "polish_klej_ner_multiple_choice"
+  },
+  "polish_dyk_regex": {
+    "exact_match,score-first": 0.8241010689990281,
+    "exact_match_stderr,score-first": 0.01187477204670372,
+    "f1,score-first": 0.6370967741935484,
+    "f1_stderr,score-first": "N/A",
+    "alias": "polish_dyk_regex"
+  },
+  "polish_dyk_multiple_choice": {
+    "acc,none": 0.8756073858114675,
+    "acc_stderr,none": 0.010293319379865268,
+    "f1,none": 0.6862745098039216,
+    "f1_stderr,none": "N/A",
+    "acc_norm,none": 0.8756073858114675,
+    "acc_norm_stderr,none": 0.010293319379865268,
+    "alias": "polish_dyk_multiple_choice"
+  },
+  "polish_belebele_regex": {
+    "exact_match,score-first": 0.8511111111111112,
+    "exact_match_stderr,score-first": 0.011872561521396008,
+    "alias": "polish_belebele_regex"
+  },
+  "polish_belebele_mc": {
+    "acc,none": 0.8166666666666667,
+    "acc_stderr,none": 0.012905156820036966,
+    "acc_norm,none": 0.8166666666666667,
+    "acc_norm_stderr,none": 0.012905156820036966,
+    "alias": "polish_belebele_mc"
+  }
+}

variant_a/eval/variant_a_gen_results.json

@@ -0,0 +1,38 @@
+{
+  "run": "Generative regex (Run 2)",
+  "model": "QuIP# E8P12 Bielik-11B-v2.3-Instruct",
+  "timestamp": "2026-02-22T00:27:00Z",
+  "total_requests": 8666,
+  "total_time_s": 10633,
+  "batch_size": 32,
+  "tasks": {
+    "polemo2_in": {
+      "exact_match,score-first": 0.8310249307479224,
+      "exact_match_stderr,score-first": 0.01395567061572903
+    },
+    "polemo2_out": {
+      "exact_match,score-first": 0.7186234817813765,
+      "exact_match_stderr,score-first": 0.020252151525388066
+    },
+    "polish_8tags_regex": {
+      "exact_match,score-first": 0.7655535224153706,
+      "exact_match_stderr,score-first": 0.0064079517407639765
+    },
+    "polish_cbd_regex": {
+      "exact_match,score-first": 0.757,
+      "exact_match_stderr,score-first": 0.013569640199177445,
+      "f1,score-first": 0.3009262592049606,
+      "f1_stderr,score-first": "N/A"
+    },
+    "polish_ppc_regex": {
+      "exact_match,score-first": 0.789,
+      "exact_match_stderr,score-first": 0.012909130321042094
+    },
+    "polish_psc_regex": {
+      "exact_match,score-first": 0.9536178107606679,
+      "exact_match_stderr,score-first": 0.0064084787268202026,
+      "f1,score-first": 0.9662162162162162,
+      "f1_stderr,score-first": "N/A"
+    }
+  }
+}

variant_a/eval/variant_a_mc_results.json

@@ -0,0 +1,52 @@
+{
+  "run": "MC loglikelihood (Run 1)",
+  "model": "QuIP# E8P12 Bielik-11B-v2.3-Instruct",
+  "timestamp": "2026-02-21T20:30:00Z",
+  "total_requests": 50996,
+  "total_time_s": 15789,
+  "tasks": {
+    "polemo2_in": {
+      "leaderboard_metric": "acc,none",
+      "leaderboard_score": 0.8518,
+      "acc": 0.8518,
+      "acc_norm": 0.1551
+    },
+    "polemo2_out": {
+      "leaderboard_metric": "acc,none",
+      "leaderboard_score": 0.7449,
+      "acc": 0.7449,
+      "acc_norm": 0.33
+    },
+    "polish_8tags": {
+      "leaderboard_metric": "acc,none",
+      "leaderboard_score": 0.7452,
+      "acc": 0.7452,
+      "acc_norm": 0.3742
+    },
+    "polish_cbd": {
+      "leaderboard_metric": "f1,none",
+      "leaderboard_score": 0.2691,
+      "acc": 0.725,
+      "acc_norm": 0.814,
+      "f1": 0.2691
+    },
+    "polish_ppc": {
+      "leaderboard_metric": "acc,none",
+      "leaderboard_score": 0.779,
+      "acc": 0.779,
+      "acc_norm": 0.779
+    },
+    "polish_psc": {
+      "leaderboard_metric": "f1,none",
+      "leaderboard_score": 0.9423,
+      "acc": 0.9657,
+      "acc_norm": 0.9657,
+      "f1": 0.9423
+    }
+  },
+  "leaderboard_avg": 0.722,
+  "baseline_iq2xxs": 0.6134,
+  "baseline_fp16": 0.6571,
+  "delta_vs_iq2xxs_pp": 10.9,
+  "delta_vs_fp16_pp": 6.5
+}

variant_a/logs/auto_chain.log

@@ -0,0 +1,123 @@
+[2026-02-21 16:08:57] Waiting for MC run (PID 9529) to finish...
+[2026-02-21 20:30:59] MC run finished!
+[2026-02-21 20:30:59] ============================================
+[2026-02-21 20:30:59]   MC RESULTS
+[2026-02-21 20:30:59] ============================================
+  polemo2_in_multiple_choice: acc_norm = 0.1551 (15.51%)
+  polemo2_out_multiple_choice: acc_norm = 0.3300 (33.00%)
+  polish_8tags_multiple_choice: acc_norm = 0.3742 (37.42%)
+  polish_cbd_multiple_choice: acc_norm = 0.8140 (81.40%)
+  polish_ppc_multiple_choice: acc_norm = 0.7790 (77.90%)
+  polish_psc_multiple_choice: acc_norm = 0.9657 (96.57%)
+
+  Average: 0.5697 (56.97%)
+  Baseline (IQ2_XXS): 61.34%
+  FP16 Instruct:      65.71%
+  >>> Below baseline by 4.37pp <<<
+[2026-02-21 20:30:59] ============================================
+[2026-02-21 20:30:59]   STARTING: Generative core tasks (6 regex)
+[2026-02-21 20:30:59] ============================================
+I0221 20:31:01.636990 11685 utils.py:148] Note: detected 192 virtual cores but NumExpr set to maximum of 64, check "NUMEXPR_MAX_THREADS" environment variable.
+I0221 20:31:01.637057 11685 utils.py:151] Note: NumExpr detected 192 cores but "NUMEXPR_MAX_THREADS" not set, so enforcing safe limit of 16.
+I0221 20:31:01.637096 11685 utils.py:164] NumExpr defaulting to 16 threads.
+I0221 20:31:01.715981 11685 config.py:58] PyTorch version 2.10.0+cu126 available.
+W0221 20:31:01.840680 11685 warnings.py:112] /dev/shm/eval/quip-sharp/lib/codebook/__init__.py:6: FutureWarning: `torch.library.impl_abstract` was renamed to `torch.library.register_fake`. Please use that instead; we will remove `torch.library.impl_abstract` in a future version of PyTorch.
+  @torch.library.impl_abstract("quip_lib::decode_matvec_e8p")
+
+W0221 20:31:01.877046 11685 warnings.py:112] /dev/shm/eval/quip-sharp/lib/codebook/__init__.py:25: FutureWarning: `torch.library.impl_abstract` was renamed to `torch.library.register_fake`. Please use that instead; we will remove `torch.library.impl_abstract` in a future version of PyTorch.
+  @torch.library.impl_abstract("quip_lib::decompress_packed_e8p")
+
+W0221 20:31:02.038564 11685 warnings.py:112] /dev/shm/eval/quip-sharp/lib/utils/matmul_had.py:96: FutureWarning: `torch.library.impl_abstract` was renamed to `torch.library.register_fake`. Please use that instead; we will remove `torch.library.impl_abstract` in a future version of PyTorch.
+  @torch.library.impl_abstract("quip_lib::hadamard")
+
+W0221 20:31:25.381577 11685 warnings.py:112] /dev/shm/eval/lm-evaluation-harness/lm_eval/filters/extraction.py:98: SyntaxWarning: invalid escape sequence '\s'
+  - step 2 : We parse the choice with regex :[\s]*([A-?]), where ? varies by number of choices.
+
+W0221 20:31:25.381833 11685 warnings.py:112] /dev/shm/eval/lm-evaluation-harness/lm_eval/filters/extraction.py:168: SyntaxWarning: invalid escape sequence '\s'
+  f":[\s]*({without_paren_fallback_regex})"
+
+[20:31:25] ============================================================
+[20:31:25]   Polish LLM Leaderboard Eval
+[20:31:25]   Model: QuIP# Bielik-Q2-Sharp Variant A
+[20:31:25]   lm-eval API: new
+[20:31:25]   Tasks: ['polemo2_in', 'polemo2_out', 'polish_8tags_regex', 'polish_cbd_regex', 'polish_ppc_regex', 'polish_psc_regex']
+[20:31:25]   Few-shot: 5
+[20:31:25] ============================================================
+[20:31:25] Loading QuIP# model from /dev/shm/eval/model...
+I0221 20:31:29.738574 11685 modeling.py:987] We will use 90% of the memory on device 0 for storing the model, and 10% for the buffer to avoid OOM. You can set `max_memory` in to a higher value to use more memory (at your own risk).
+[20:31:31] Model loaded in 5.8s
+[20:31:31] Tokenizer: speakleash/Bielik-11B-v2.3-Instruct (vocab=32000)
+[20:31:31] 
+Running 6 tasks...
+I0221 20:31:31.609011 11685 evaluator.py:152] Setting random seed to 0 | Setting numpy seed to 1234 | Setting torch manual seed to 1234
+I0221 20:31:31.609088 11685 evaluator.py:203] Using pre-initialized model
+W0221 20:31:45.882049 11685 evaluator.py:251] Overwriting default num_fewshot of polish_psc_regex from None to 5
+I0221 20:31:45.882143 11685 evaluator.py:261] Setting fewshot random generator seed to 1234
+W0221 20:31:45.882179 11685 evaluator.py:251] Overwriting default num_fewshot of polish_ppc_regex from None to 5
+I0221 20:31:45.882212 11685 evaluator.py:261] Setting fewshot random generator seed to 1234
+W0221 20:31:45.882230 11685 evaluator.py:251] Overwriting default num_fewshot of polish_cbd_regex from None to 5
+I0221 20:31:45.882256 11685 evaluator.py:261] Setting fewshot random generator seed to 1234
+W0221 20:31:45.882269 11685 evaluator.py:251] Overwriting default num_fewshot of polish_8tags_regex from None to 5
+I0221 20:31:45.882290 11685 evaluator.py:261] Setting fewshot random generator seed to 1234
+W0221 20:31:45.882302 11685 evaluator.py:251] Overwriting default num_fewshot of polemo2_out from None to 5
+I0221 20:31:45.882324 11685 evaluator.py:261] Setting fewshot random generator seed to 1234
+W0221 20:31:45.882337 11685 evaluator.py:251] Overwriting default num_fewshot of polemo2_in from None to 5
+I0221 20:31:45.882368 11685 evaluator.py:261] Setting fewshot random generator seed to 1234
+I0221 20:31:45.883396 11685 task.py:410] Building contexts for polish_psc_regex on rank 0...
+
  0%|          | 0/1078 [00:00<?, ?it/s]
  2%|▏         | 17/1078 [00:00<00:06, 153.98it/s]
  3%|▎         | 33/1078 [00:00<00:06, 156.96it/s]
  5%|▍         | 50/1078 [00:00<00:06, 158.59it/s]
  6%|▌         | 66/1078 [00:00<00:06, 156.64it/s]
  8%|▊         | 83/1078 [00:00<00:06, 158.02it/s]
  9%|▉         | 99/1078 [00:00<00:06, 156.52it/s]
 11%|█         | 116/1078 [00:00<00:06, 157.87it/s]
 12%|█▏        | 132/1078 [00:00<00:06, 156.62it/s]
 14%|█▍        | 149/1078 [00:00<00:05, 157.99it/s]
 15%|█▌        | 166/1078 [00:01<00:05, 156.89it/s]
 17%|█▋        | 183/1078 [00:01<00:05, 158.16it/s]
 19%|█▊        | 200/1078 [00:01<00:05, 159.00it/s]
 20%|██        | 216/1078 [00:01<00:05, 158.96it/s]
 22%|██▏       | 232/1078 [00:01<00:05, 158.60it/s]
 23%|██▎       | 249/1078 [00:01<00:05, 159.21it/s]
 25%|██▍       | 265/1078 [00:01<00:05, 158.80it/s]
 26%|██▌       | 282/1078 [00:01<00:04, 159.40it/s]
 28%|██▊       | 299/1078 [00:01<00:04, 159.87it/s]
 29%|██▉       | 316/1078 [00:01<00:04, 160.17it/s]
 31%|███       | 333/1078 [00:02<00:04, 159.55it/s]
 32%|███▏      | 350/1078 [00:02<00:04, 159.86it/s]
 34%|███▍      | 367/1078 [00:02<00:04, 160.39it/s]
 36%|███▌      | 384/1078 [00:02<00:04, 160.57it/s]
 37%|███▋      | 401/1078 [00:02<00:04, 159.93it/s]
 39%|███▊      | 417/1078 [00:02<00:04, 159.13it/s]
 40%|████      | 434/1078 [00:02<00:04, 159.49it/s]
 42%|████▏     | 451/1078 [00:02<00:03, 159.98it/s]
 43%|████▎     | 468/1078 [00:02<00:03, 160.05it/s]
 45%|████▍     | 485/1078 [00:03<00:03, 160.08it/s]
 47%|████▋     | 502/1078 [00:03<00:03, 160.08it/s]
 48%|████▊     | 519/1078 [00:03<00:03, 160.03it/s]
 50%|████▉     | 536/1078 [00:03<00:03, 158.67it/s]
 51%|█████▏    | 553/1078 [00:03<00:03, 159.22it/s]
 53%|█████▎    | 569/1078 [00:03<00:03, 159.38it/s]
 54%|█████▍    | 585/1078 [00:03<00:03, 159.53it/s]
 56%|█████▌    | 601/1078 [00:03<00:02, 159.21it/s]
 57%|█████▋    | 617/1078 [00:03<00:02, 159.38it/s]
 59%|█████▉    | 634/1078 [00:03<00:02, 159.72it/s]
 60%|██████    | 650/1078 [00:04<00:02, 159.65it/s]
 62%|██████▏   | 666/1078 [00:04<00:02, 159.52it/s]
 63%|██████▎   | 683/1078 [00:04<00:02, 159.71it/s]
 65%|██████▍   | 699/1078 [00:04<00:02, 159.74it/s]
 66%|██████▋   | 715/1078 [00:04<00:02, 159.73it/s]
 68%|██████▊   | 732/1078 [00:04<00:02, 160.31it/s]
 69%|██████▉   | 749/1078 [00:04<00:02, 160.44it/s]
 71%|███████   | 766/1078 [00:04<00:01, 160.44it/s]
 73%|███████▎  | 783/1078 [00:04<00:01, 160.81it/s]
 74%|███████▍  | 800/1078 [00:05<00:01, 160.64it/s]
 76%|███████▌  | 817/1078 [00:05<00:01, 160.53it/s]
 77%|███████▋  | 834/1078 [00:05<00:01, 160.69it/s]
 79%|███████▉  | 851/1078 [00:05<00:01, 160.49it/s]
 81%|████████  | 868/1078 [00:05<00:01, 160.85it/s]
 82%|████████▏ | 885/1078 [00:05<00:01, 160.85it/s]
 84%|████████▎ | 902/1078 [00:05<00:01, 156.37it/s]
 85%|████████▌ | 918/1078 [00:05<00:01, 156.26it/s]
 87%|████████▋ | 934/1078 [00:05<00:00, 156.76it/s]
 88%|████████▊ | 951/1078 [00:05<00:00, 158.04it/s]
 90%|████████▉ | 968/1078 [00:06<00:00, 158.99it/s]
 91%|█████████▏| 985/1078 [00:06<00:00, 159.38it/s]
 93%|█████████▎| 1001/1078 [00:06<00:00, 159.27it/s]
 94%|█████████▍| 1018/1078 [00:06<00:00, 159.91it/s]
 96%|█████████▌| 1035/1078 [00:06<00:00, 160.10it/s]
 98%|█████████▊| 1052/1078 [00:06<00:00, 160.18it/s]
 99%|█████████▉| 1069/1078 [00:06<00:00, 160.22it/s]
+I0221 20:31:52.669470 11685 task.py:410] Building contexts for polish_ppc_regex on rank 0...
+
  0%|          | 0/1000 [00:00<?, ?it/s]
  2%|▏         | 17/1000 [00:00<00:05, 169.98it/s]
  4%|▎         | 35/1000 [00:00<00:05, 170.34it/s]
  5%|▌         | 53/1000 [00:00<00:05, 170.27it/s]
  7%|▋         | 71/1000 [00:00<00:05, 169.37it/s]
  9%|▉         | 89/1000 [00:00<00:05, 169.88it/s]
 11%|█         | 107/1000 [00:00<00:05, 169.72it/s]
 12%|█▎        | 125/1000 [00:00<00:05, 170.09it/s]
 14%|█▍        | 143/1000 [00:00<00:05, 170.19it/s]
 16%|█▌        | 161/1000 [00:00<00:04, 170.35it/s]
 18%|█▊        | 179/1000 [00:01<00:04, 169.73it/s]
 20%|█▉        | 197/1000 [00:01<00:04, 170.16it/s]
 22%|██▏       | 215/1000 [00:01<00:04, 170.28it/s]
 23%|██▎       | 233/1000 [00:01<00:04, 170.40it/s]
 25%|██▌       | 251/1000 [00:01<00:04, 170.31it/s]
 27%|██▋       | 269/1000 [00:01<00:04, 170.45it/s]
 29%|██▊       | 287/1000 [00:01<00:04, 169.65it/s]
 30%|███       | 305/1000 [00:01<00:04, 169.83it/s]
 32%|███▏      | 322/1000 [00:01<00:04, 169.43it/s]
 34%|███▍      | 339/1000 [00:01<00:03, 169.05it/s]
 36%|███▌      | 356/1000 [00:02<00:04, 160.66it/s]
 37%|███▋      | 373/1000 [00:02<00:03, 156.82it/s]
 39%|███▉      | 389/1000 [00:02<00:03, 157.01it/s]
 40%|████      | 405/1000 [00:02<00:04, 148.01it/s]
 42%|████▏     | 420/1000 [00:02<00:03, 148.03it/s]
 44%|████▎     | 437/1000 [00:02<00:03, 152.61it/s]
 45%|████▌     | 454/1000 [00:02<00:03, 157.56it/s]
 47%|████▋     | 472/1000 [00:02<00:03, 161.56it/s]
 49%|████▉     | 490/1000 [00:02<00:03, 164.47it/s]
 51%|█████     | 508/1000 [00:03<00:02, 166.24it/s]
 53%|█████▎    | 526/1000 [00:03<00:02, 167.75it/s]
 54%|█████▍    | 543/1000 [00:03<00:02, 167.98it/s]
 56%|█████▌    | 561/1000 [00:03<00:02, 168.75it/s]
 58%|█████▊    | 579/1000 [00:03<00:02, 169.26it/s]
 60%|█████▉    | 597/1000 [00:03<00:02, 169.86it/s]
 61%|██████▏   | 614/1000 [00:03<00:02, 164.36it/s]
 63%|██████▎   | 631/1000 [00:03<00:02, 165.89it/s]
 65%|██████▍   | 649/1000 [00:03<00:02, 167.29it/s]
 67%|██████▋   | 667/1000 [00:04<00:01, 168.11it/s]
 68%|██████▊   | 685/1000 [00:04<00:01, 168.78it/s]
 70%|███████   | 703/1000 [00:04<00:01, 169.39it/s]
 72%|███████▏  | 720/1000 [00:04<00:01, 168.94it/s]
 74%|███████▎  | 737/1000 [00:04<00:01, 168.72it/s]
 76%|███████▌  | 755/1000 [00:04<00:01, 169.38it/s]
 77%|███████▋  | 773/1000 [00:04<00:01, 169.79it/s]
 79%|███████▉  | 791/1000 [00:04<00:01, 170.04it/s]
 81%|████████  | 809/1000 [00:04<00:01, 170.21it/s]
 83%|████████▎ | 827/1000 [00:04<00:01, 170.34it/s]
 84%|████████▍ | 845/1000 [00:05<00:00, 170.43it/s]
 86%|████████▋ | 863/1000 [00:05<00:00, 170.74it/s]
 88%|████████▊ | 881/1000 [00:05<00:00, 170.93it/s]
 90%|████████▉ | 899/1000 [00:05<00:00, 171.18it/s]
 92%|█████████▏| 917/1000 [00:05<00:00, 171.08it/s]
 94%|█████████▎| 935/1000 [00:05<00:00, 171.23it/s]
 95%|█████████▌| 953/1000 [00:05<00:00, 171.30it/s]
 97%|█████████▋| 971/1000 [00:05<00:00, 171.48it/s]
 99%|█████████▉| 989/1000 [00:05<00:00, 171.54it/s]
+I0221 20:31:58.657753 11685 task.py:410] Building contexts for polish_cbd_regex on rank 0...
+
  0%|          | 0/1000 [00:00<?, ?it/s]
  2%|▏         | 19/1000 [00:00<00:05, 189.32it/s]
  4%|▍         | 38/1000 [00:00<00:05, 189.43it/s]
  6%|▌         | 57/1000 [00:00<00:04, 189.66it/s]
  8%|▊         | 77/1000 [00:00<00:04, 190.06it/s]
 10%|▉         | 97/1000 [00:00<00:04, 189.98it/s]
 12%|█▏        | 116/1000 [00:00<00:04, 189.98it/s]
 14%|█▎        | 136/1000 [00:00<00:04, 190.05it/s]
 16%|█▌        | 156/1000 [00:00<00:04, 190.09it/s]
 18%|█▊        | 176/1000 [00:00<00:04, 190.11it/s]
 20%|█▉        | 196/1000 [00:01<00:04, 190.07it/s]
 22%|██▏       | 216/1000 [00:01<00:04, 190.29it/s]
 24%|██▎       | 236/1000 [00:01<00:04, 190.44it/s]
 26%|██▌       | 256/1000 [00:01<00:03, 190.86it/s]
 28%|██▊       | 276/1000 [00:01<00:03, 190.71it/s]
 30%|██▉       | 296/1000 [00:01<00:03, 189.08it/s]
 32%|███▏      | 315/1000 [00:01<00:03, 189.21it/s]
 34%|███▎      | 335/1000 [00:01<00:03, 189.53it/s]
 36%|███▌      | 355/1000 [00:01<00:03, 189.89it/s]
 37%|███▋      | 374/1000 [00:01<00:03, 189.69it/s]
 39%|███▉      | 393/1000 [00:02<00:03, 189.23it/s]
 41%|████▏     | 413/1000 [00:02<00:03, 189.48it/s]
 43%|████▎     | 432/1000 [00:02<00:02, 189.43it/s]
 45%|████▌     | 452/1000 [00:02<00:02, 189.66it/s]
 47%|████▋     | 471/1000 [00:02<00:02, 189.76it/s]
 49%|████▉     | 491/1000 [00:02<00:02, 189.63it/s]
 51%|█████     | 511/1000 [00:02<00:02, 189.75it/s]
 53%|█████▎    | 531/1000 [00:02<00:02, 189.97it/s]
 55%|█████▌    | 551/1000 [00:02<00:02, 190.07it/s]
 57%|█████▋    | 571/1000 [00:03<00:02, 190.07it/s]
 59%|█████▉    | 591/1000 [00:03<00:02, 189.98it/s]
 61%|██████    | 610/1000 [00:03<00:02, 189.24it/s]
 63%|██████▎   | 629/1000 [00:03<00:01, 189.01it/s]
 65%|██████▍   | 648/1000 [00:03<00:01, 189.26it/s]
 67%|██████▋   | 667/1000 [00:03<00:01, 189.43it/s]
 69%|██████▊   | 686/1000 [00:03<00:01, 189.30it/s]
 71%|███████   | 706/1000 [00:03<00:01, 189.63it/s]
 73%|███████▎  | 726/1000 [00:03<00:01, 189.79it/s]
 74%|███████▍  | 745/1000 [00:03<00:01, 189.57it/s]
 76%|███████▋  | 765/1000 [00:04<00:01, 189.81it/s]
 78%|███████▊  | 785/1000 [00:04<00:01, 189.98it/s]
 80%|████████  | 805/1000 [00:04<00:01, 190.01it/s]
 82%|████████▎ | 825/1000 [00:04<00:00, 190.37it/s]
 84%|████████▍ | 845/1000 [00:04<00:00, 190.52it/s]
 86%|████████▋ | 865/1000 [00:04<00:00, 190.52it/s]
 88%|████████▊ | 885/1000 [00:04<00:00, 190.55it/s]
 90%|█████████ | 905/1000 [00:04<00:00, 190.66it/s]
 92%|█████████▎| 925/1000 [00:04<00:00, 190.64it/s]
 94%|█████████▍| 945/1000 [00:04<00:00, 190.52it/s]
 96%|█████████▋| 965/1000 [00:05<00:00, 190.90it/s]
 98%|█████████▊| 985/1000 [00:05<00:00, 187.92it/s]
+I0221 20:32:03.950957 11685 task.py:410] Building contexts for polish_8tags_regex on rank 0...
+
  0%|          | 0/4372 [00:00<?, ?it/s]
  0%|          | 17/4372 [00:00<00:26, 165.37it/s]
  1%|          | 35/4372 [00:00<00:25, 169.45it/s]
  1%|          | 53/4372 [00:00<00:25, 170.00it/s]
  2%|▏         | 71/4372 [00:00<00:25, 170.98it/s]
  2%|▏         | 89/4372 [00:00<00:24, 171.57it/s]
  2%|▏         | 107/4372 [00:00<00:24, 171.56it/s]
  3%|▎         | 125/4372 [00:00<00:25, 166.06it/s]
  3%|▎         | 142/4372 [00:00<00:26, 157.02it/s]
  4%|▎         | 158/4372 [00:00<00:26, 156.58it/s]
  4%|▍         | 175/4372 [00:01<00:26, 158.31it/s]
  4%|▍         | 193/4372 [00:01<00:25, 162.59it/s]
  5%|▍         | 211/4372 [00:01<00:25, 165.06it/s]
  5%|▌         | 229/4372 [00:01<00:24, 167.21it/s]
  6%|▌         | 247/4372 [00:01<00:24, 168.54it/s]
  6%|▌         | 265/4372 [00:01<00:24, 169.67it/s]
  6%|▋         | 283/4372 [00:01<00:24, 170.05it/s]
  7%|▋         | 301/4372 [00:01<00:23, 170.59it/s]
  7%|▋         | 319/4372 [00:01<00:23, 170.99it/s]
  8%|▊         | 337/4372 [00:02<00:23, 171.05it/s]
  8%|▊         | 355/4372 [00:02<00:23, 171.32it/s]
  9%|▊         | 373/4372 [00:02<00:23, 171.60it/s]
  9%|▉         | 391/4372 [00:02<00:23, 171.96it/s]
  9%|▉         | 409/4372 [00:02<00:23, 172.20it/s]
 10%|▉         | 427/4372 [00:02<00:22, 172.27it/s]
 10%|█         | 445/4372 [00:02<00:22, 172.51it/s]
 11%|█         | 463/4372 [00:02<00:22, 172.50it/s]
 11%|█         | 481/4372 [00:02<00:22, 172.69it/s]
 11%|█▏        | 499/4372 [00:02<00:22, 172.60it/s]
 12%|█▏        | 517/4372 [00:03<00:22, 172.65it/s]
 12%|█▏        | 535/4372 [00:03<00:22, 172.62it/s]
 13%|█▎        | 553/4372 [00:03<00:22, 172.58it/s]
 13%|█▎        | 571/4372 [00:03<00:22, 172.41it/s]
 13%|█▎        | 589/4372 [00:03<00:21, 172.41it/s]
 14%|█▍        | 607/4372 [00:03<00:21, 172.52it/s]
 14%|█▍        | 625/4372 [00:03<00:21, 172.68it/s]
 15%|█▍        | 643/4372 [00:03<00:21, 172.67it/s]
 15%|█▌        | 661/4372 [00:03<00:21, 172.78it/s]
 16%|█▌        | 679/4372 [00:03<00:21, 172.62it/s]
 16%|█▌        | 697/4372 [00:04<00:21, 172.35it/s]
 16%|█▋        | 715/4372 [00:04<00:21, 172.50it/s]
 17%|█▋        | 733/4372 [00:04<00:21, 172.51it/s]
 17%|█▋        | 751/4372 [00:04<00:21, 172.37it/s]
 18%|█▊        | 769/4372 [00:04<00:20, 172.09it/s]
 18%|█▊        | 787/4372 [00:04<00:20, 172.06it/s]
 18%|█▊        | 805/4372 [00:04<00:20, 171.80it/s]
 19%|█▉        | 823/4372 [00:04<00:20, 171.87it/s]
 19%|█▉        | 841/4372 [00:04<00:20, 171.86it/s]
 20%|█▉        | 859/4372 [00:05<00:20, 172.00it/s]
 20%|██        | 877/4372 [00:05<00:20, 171.85it/s]
 20%|██        | 895/4372 [00:05<00:20, 172.01it/s]
 21%|██        | 913/4372 [00:05<00:20, 171.39it/s]
 21%|██▏       | 931/4372 [00:05<00:20, 170.58it/s]
 22%|██▏       | 949/4372 [00:05<00:20, 170.34it/s]
 22%|██▏       | 967/4372 [00:05<00:19, 170.60it/s]
 23%|██▎       | 985/4372 [00:05<00:19, 170.82it/s]
 23%|██▎       | 1003/4372 [00:05<00:19, 171.03it/s]
 23%|██▎       | 1021/4372 [00:05<00:19, 170.36it/s]
 24%|██▍       | 1039/4372 [00:06<00:19, 170.87it/s]
 24%|██▍       | 1057/4372 [00:06<00:19, 171.29it/s]
 25%|██▍       | 1075/4372 [00:06<00:19, 171.64it/s]
 25%|██▌       | 1093/4372 [00:06<00:19, 171.16it/s]
 25%|██▌       | 1111/4372 [00:06<00:19, 171.34it/s]
 26%|██▌       | 1129/4372 [00:06<00:18, 171.35it/s]
 26%|██▌       | 1147/4372 [00:06<00:18, 171.40it/s]
 27%|██▋       | 1165/4372 [00:06<00:18, 170.93it/s]
 27%|██▋       | 1183/4372 [00:06<00:18, 171.06it/s]
 27%|██▋       | 1201/4372 [00:07<00:18, 171.31it/s]
 28%|██▊       | 1219/4372 [00:07<00:18, 171.24it/s]
 28%|██▊       | 1237/4372 [00:07<00:18, 171.38it/s]
 29%|██▊       | 1255/4372 [00:07<00:18, 171.11it/s]
 29%|██▉       | 1273/4372 [00:07<00:18, 166.70it/s]
 30%|██▉       | 1290/4372 [00:07<00:18, 165.22it/s]
 30%|██▉       | 1307/4372 [00:07<00:18, 165.70it/s]
 30%|███       | 1325/4372 [00:07<00:18, 167.27it/s]
 31%|███       | 1343/4372 [00:07<00:17, 168.41it/s]
 31%|███       | 1361/4372 [00:07<00:17, 169.07it/s]
 32%|███▏      | 1379/4372 [00:08<00:17, 169.54it/s]
 32%|███▏      | 1397/4372 [00:08<00:17, 170.01it/s]
 32%|███▏      | 1415/4372 [00:08<00:17, 170.14it/s]
 33%|███▎      | 1433/4372 [00:08<00:17, 170.45it/s]
 33%|███▎      | 1451/4372 [00:08<00:17, 170.38it/s]
 34%|███▎      | 1469/4372 [00:08<00:17, 170.62it/s]
 34%|███▍      | 1487/4372 [00:08<00:16, 170.71it/s]
 34%|███▍      | 1505/4372 [00:08<00:16, 168.73it/s]
 35%|███▍      | 1522/4372 [00:08<00:16, 168.72it/s]
 35%|███▌      | 1539/4372 [00:09<00:16, 168.70it/s]
 36%|███▌      | 1556/4372 [00:09<00:16, 167.92it/s]
 36%|███▌      | 1574/4372 [00:09<00:16, 168.80it/s]
 36%|███▋      | 1591/4372 [00:09<00:16, 168.43it/s]
 37%|███▋      | 1608/4372 [00:09<00:17, 159.94it/s]
 37%|███▋      | 1625/4372 [00:09<00:18, 151.78it/s]
 38%|███▊      | 1641/4372 [00:09<00:18, 150.90it/s]
 38%|███▊      | 1657/4372 [00:09<00:18, 148.23it/s]
 38%|███▊      | 1672/4372 [00:09<00:18, 145.93it/s]
 39%|███▊      | 1688/4372 [00:10<00:18, 147.94it/s]
 39%|███▉      | 1705/4372 [00:10<00:17, 153.33it/s]
 39%|███▉      | 1723/4372 [00:10<00:16, 158.39it/s]
 40%|███▉      | 1741/4372 [00:10<00:16, 162.06it/s]
 40%|████      | 1759/4372 [00:10<00:15, 164.68it/s]
 41%|████      | 1776/4372 [00:10<00:15, 166.14it/s]
 41%|████      | 1794/4372 [00:10<00:15, 167.72it/s]
 41%|████▏     | 1812/4372 [00:10<00:15, 169.29it/s]
 42%|████▏     | 1830/4372 [00:10<00:14, 169.82it/s]
 42%|████▏     | 1848/4372 [00:10<00:14, 169.87it/s]
 43%|████▎     | 1866/4372 [00:11<00:14, 170.57it/s]
 43%|████▎     | 1884/4372 [00:11<00:14, 167.64it/s]
 44%|████▎     | 1902/4372 [00:11<00:14, 169.00it/s]
 44%|████▍     | 1920/4372 [00:11<00:14, 169.80it/s]
 44%|████▍     | 1938/4372 [00:11<00:14, 170.60it/s]
 45%|████▍     | 1956/4372 [00:11<00:14, 171.06it/s]
 45%|████▌     | 1974/4372 [00:11<00:13, 171.60it/s]
 46%|████▌     | 1992/4372 [00:11<00:13, 171.95it/s]
 46%|████▌     | 2010/4372 [00:11<00:13, 172.22it/s]
 46%|████▋     | 2028/4372 [00:12<00:13, 172.41it/s]
 47%|████▋     | 2046/4372 [00:12<00:13, 172.88it/s]
 47%|████▋     | 2064/4372 [00:12<00:13, 172.70it/s]
 48%|████▊     | 2082/4372 [00:12<00:13, 172.76it/s]
 48%|████▊     | 2100/4372 [00:12<00:13, 172.77it/s]
 48%|████▊     | 2118/4372 [00:12<00:13, 172.90it/s]
 49%|████▉     | 2136/4372 [00:12<00:12, 173.07it/s]
 49%|████▉     | 2154/4372 [00:12<00:12, 172.29it/s]
 50%|████▉     | 2172/4372 [00:12<00:12, 172.33it/s]
 50%|█████     | 2190/4372 [00:12<00:12, 171.54it/s]
 51%|█████     | 2208/4372 [00:13<00:12, 171.31it/s]
 51%|█████     | 2226/4372 [00:13<00:12, 170.73it/s]
 51%|█████▏    | 2244/4372 [00:13<00:12, 169.71it/s]
 52%|█████▏    | 2262/4372 [00:13<00:12, 170.28it/s]
 52%|█████▏    | 2280/4372 [00:13<00:12, 170.65it/s]
 53%|█████▎    | 2298/4372 [00:13<00:12, 171.15it/s]
 53%|█████▎    | 2316/4372 [00:13<00:11, 171.53it/s]
 53%|█████▎    | 2334/4372 [00:13<00:11, 171.78it/s]
 54%|█████▍    | 2352/4372 [00:13<00:11, 172.01it/s]
 54%|█████▍    | 2370/4372 [00:14<00:11, 172.32it/s]
 55%|█████▍    | 2388/4372 [00:14<00:11, 172.28it/s]
 55%|█████▌    | 2406/4372 [00:14<00:11, 172.63it/s]
 55%|█████▌    | 2424/4372 [00:14<00:11, 172.40it/s]
 56%|█████▌    | 2442/4372 [00:14<00:11, 172.60it/s]
 56%|█████▋    | 2460/4372 [00:14<00:11, 172.34it/s]
 57%|█████▋    | 2478/4372 [00:14<00:10, 172.40it/s]
 57%|█████▋    | 2496/4372 [00:14<00:10, 172.25it/s]
 58%|█████▊    | 2514/4372 [00:14<00:10, 172.08it/s]
 58%|█████▊    | 2532/4372 [00:14<00:10, 172.02it/s]
 58%|█████▊    | 2550/4372 [00:15<00:10, 172.03it/s]
 59%|█████▊    | 2568/4372 [00:15<00:10, 171.97it/s]
 59%|█████▉    | 2586/4372 [00:15<00:10, 171.46it/s]
 60%|█████▉    | 2604/4372 [00:15<00:10, 169.89it/s]
 60%|█████▉    | 2622/4372 [00:15<00:10, 170.36it/s]
 60%|██████    | 2640/4372 [00:15<00:10, 170.57it/s]
 61%|██████    | 2658/4372 [00:15<00:10, 171.06it/s]
 61%|██████    | 2676/4372 [00:15<00:09, 170.85it/s]
 62%|██████▏   | 2694/4372 [00:15<00:09, 171.28it/s]
 62%|██████▏   | 2712/4372 [00:16<00:09, 170.94it/s]
 62%|██████▏   | 2730/4372 [00:16<00:09, 171.27it/s]
 63%|██████▎   | 2748/4372 [00:16<00:09, 171.62it/s]
 63%|██████▎   | 2766/4372 [00:16<00:09, 171.38it/s]
 64%|██████▎   | 2784/4372 [00:16<00:09, 171.82it/s]
 64%|██████▍   | 2802/4372 [00:16<00:09, 171.84it/s]
 65%|██████▍   | 2820/4372 [00:16<00:09, 172.27it/s]
 65%|██████▍   | 2838/4372 [00:16<00:08, 172.39it/s]
 65%|██████▌   | 2856/4372 [00:16<00:08, 172.41it/s]
 66%|██████▌   | 2874/4372 [00:16<00:08, 172.79it/s]
 66%|██████▌   | 2892/4372 [00:17<00:08, 172.97it/s]
 67%|██████▋   | 2910/4372 [00:17<00:08, 172.51it/s]
 67%|██████▋   | 2928/4372 [00:17<00:08, 171.73it/s]
 67%|██████▋   | 2946/4372 [00:17<00:08, 171.92it/s]
 68%|██████▊   | 2964/4372 [00:17<00:08, 172.03it/s]
 68%|██████▊   | 2982/4372 [00:17<00:08, 172.29it/s]
 69%|██████▊   | 3000/4372 [00:17<00:07, 172.34it/s]
 69%|██████▉   | 3018/4372 [00:17<00:07, 172.46it/s]
 69%|██████▉   | 3036/4372 [00:17<00:07, 172.44it/s]
 70%|██████▉   | 3054/4372 [00:17<00:07, 172.62it/s]
 70%|███████   | 3072/4372 [00:18<00:07, 172.97it/s]
 71%|███████   | 3090/4372 [00:18<00:07, 172.81it/s]
 71%|███████   | 3108/4372 [00:18<00:07, 173.00it/s]
 72%|███████▏  | 3126/4372 [00:18<00:07, 173.28it/s]
 72%|███████▏  | 3144/4372 [00:18<00:07, 172.18it/s]
 72%|███████▏  | 3162/4372 [00:18<00:07, 172.69it/s]
 73%|███████▎  | 3180/4372 [00:18<00:06, 172.03it/s]
 73%|███████▎  | 3198/4372 [00:18<00:06, 172.41it/s]
 74%|███████▎  | 3216/4372 [00:18<00:06, 172.49it/s]
 74%|███████▍  | 3234/4372 [00:19<00:06, 172.39it/s]
 74%|███████▍  | 3252/4372 [00:19<00:06, 172.87it/s]
 75%|███████▍  | 3270/4372 [00:19<00:06, 172.87it/s]
 75%|███████▌  | 3288/4372 [00:19<00:06, 172.18it/s]
 76%|███████▌  | 3306/4372 [00:19<00:06, 172.34it/s]
 76%|███████▌  | 3324/4372 [00:19<00:06, 172.05it/s]
 76%|███████▋  | 3342/4372 [00:19<00:05, 172.30it/s]
 77%|███████▋  | 3360/4372 [00:19<00:05, 172.45it/s]
 77%|███████▋  | 3378/4372 [00:19<00:05, 172.86it/s]
 78%|███████▊  | 3396/4372 [00:19<00:05, 172.66it/s]
 78%|███████▊  | 3414/4372 [00:20<00:05, 172.93it/s]
 78%|███████▊  | 3432/4372 [00:20<00:05, 173.17it/s]
 79%|███████▉  | 3450/4372 [00:20<00:05, 173.37it/s]
 79%|███████▉  | 3468/4372 [00:20<00:05, 172.77it/s]
 80%|███████▉  | 3486/4372 [00:20<00:05, 172.84it/s]
 80%|████████  | 3504/4372 [00:20<00:05, 172.62it/s]
 81%|████████  | 3522/4372 [00:20<00:04, 172.76it/s]
 81%|████████  | 3540/4372 [00:20<00:04, 172.98it/s]
 81%|████████▏ | 3558/4372 [00:20<00:04, 172.85it/s]
 82%|████████▏ | 3576/4372 [00:21<00:04, 172.84it/s]
 82%|████████▏ | 3594/4372 [00:21<00:04, 172.85it/s]
 83%|████████▎ | 3612/4372 [00:21<00:04, 172.81it/s]
 83%|████████▎ | 3630/4372 [00:21<00:04, 172.72it/s]
 83%|████████▎ | 3648/4372 [00:21<00:04, 172.72it/s]
 84%|████████▍ | 3666/4372 [00:21<00:04, 173.01it/s]
 84%|████████▍ | 3684/4372 [00:21<00:03, 173.13it/s]
 85%|████████▍ | 3702/4372 [00:21<00:03, 173.16it/s]
 85%|████████▌ | 3720/4372 [00:21<00:03, 173.33it/s]
 85%|████████▌ | 3738/4372 [00:21<00:03, 173.12it/s]
 86%|████████▌ | 3756/4372 [00:22<00:03, 173.13it/s]
 86%|████████▋ | 3774/4372 [00:22<00:03, 172.98it/s]
 87%|████████▋ | 3792/4372 [00:22<00:03, 173.35it/s]
 87%|████████▋ | 3810/4372 [00:22<00:03, 173.45it/s]
 88%|████████▊ | 3828/4372 [00:22<00:03, 173.43it/s]
 88%|████████▊ | 3846/4372 [00:22<00:03, 173.62it/s]
 88%|████████▊ | 3864/4372 [00:22<00:02, 173.51it/s]
 89%|████████▉ | 3882/4372 [00:22<00:02, 173.53it/s]
 89%|████████▉ | 3900/4372 [00:22<00:02, 173.53it/s]
 90%|████████▉ | 3918/4372 [00:22<00:02, 173.19it/s]
 90%|█████████ | 3936/4372 [00:23<00:02, 172.62it/s]
 90%|█████████ | 3954/4372 [00:23<00:02, 172.83it/s]
 91%|█████████ | 3972/4372 [00:23<00:02, 172.87it/s]
 91%|█████████▏| 3990/4372 [00:23<00:02, 172.96it/s]
 92%|█████████▏| 4008/4372 [00:23<00:02, 173.00it/s]
 92%|█████████▏| 4026/4372 [00:23<00:01, 173.17it/s]
 92%|█████████▏| 4044/4372 [00:23<00:01, 173.37it/s]
 93%|█████████▎| 4062/4372 [00:23<00:01, 173.08it/s]
 93%|█████████▎| 4080/4372 [00:23<00:01, 172.68it/s]
 94%|█████████▎| 4098/4372 [00:24<00:01, 171.68it/s]
 94%|█████████▍| 4116/4372 [00:24<00:01, 171.09it/s]
 95%|█████████▍| 4134/4372 [00:24<00:01, 171.19it/s]
 95%|█████████▍| 4152/4372 [00:24<00:01, 171.88it/s]
 95%|█████████▌| 4170/4372 [00:24<00:01, 172.64it/s]
 96%|█████████▌| 4188/4372 [00:24<00:01, 173.08it/s]
 96%|█████████▌| 4206/4372 [00:24<00:00, 172.97it/s]
 97%|█████████▋| 4224/4372 [00:24<00:00, 173.12it/s]
 97%|█████████▋| 4242/4372 [00:24<00:00, 173.31it/s]
 97%|█████████▋| 4260/4372 [00:24<00:00, 173.51it/s]
 98%|█████████▊| 4278/4372 [00:25<00:00, 173.39it/s]
 98%|█████████▊| 4296/4372 [00:25<00:00, 173.63it/s]
 99%|█████████▊| 4314/4372 [00:25<00:00, 173.70it/s]
 99%|█████████▉| 4332/4372 [00:25<00:00, 173.80it/s]
 99%|█████████▉| 4350/4372 [00:25<00:00, 173.57it/s]
+I0221 20:32:29.614308 11685 task.py:410] Building contexts for polemo2_out on rank 0...
+
  0%|          | 0/494 [00:00<?, ?it/s]
  4%|▍         | 21/494 [00:00<00:02, 203.65it/s]
  9%|▊         | 42/494 [00:00<00:02, 204.62it/s]
 13%|█▎        | 63/494 [00:00<00:02, 204.95it/s]
 17%|█▋        | 84/494 [00:00<00:01, 205.27it/s]
 21%|██▏       | 105/494 [00:00<00:01, 205.68it/s]
 26%|██▌       | 126/494 [00:00<00:01, 205.86it/s]
 30%|██▉       | 147/494 [00:00<00:01, 205.88it/s]
 34%|███▍      | 168/494 [00:00<00:01, 206.00it/s]
 38%|███▊      | 189/494 [00:00<00:01, 206.13it/s]
 43%|████▎     | 210/494 [00:01<00:01, 206.23it/s]
 47%|████▋     | 231/494 [00:01<00:01, 205.99it/s]
 51%|█████     | 252/494 [00:01<00:01, 205.83it/s]
 55%|█████▌    | 273/494 [00:01<00:01, 206.00it/s]
 60%|█████▉    | 294/494 [00:01<00:00, 206.01it/s]
 64%|██████▍   | 315/494 [00:01<00:00, 205.92it/s]
 68%|██████▊   | 336/494 [00:01<00:00, 206.10it/s]
 72%|███████▏  | 357/494 [00:02<00:01, 91.16it/s] 
 77%|███████▋  | 378/494 [00:02<00:01, 109.18it/s]
 81%|████████  | 398/494 [00:02<00:00, 125.68it/s]
 85%|████████▍ | 419/494 [00:02<00:00, 141.91it/s]
 89%|████████▉ | 440/494 [00:02<00:00, 155.90it/s]
 93%|█████████▎| 461/494 [00:02<00:00, 167.43it/s]
 97%|█████████▋| 481/494 [00:02<00:00, 175.44it/s]
+I0221 20:32:32.467047 11685 task.py:410] Building contexts for polemo2_in on rank 0...
+
  0%|          | 0/722 [00:00<?, ?it/s]
  3%|▎         | 20/722 [00:00<00:03, 199.55it/s]
  6%|▌         | 41/722 [00:00<00:03, 200.29it/s]
  9%|▊         | 62/722 [00:00<00:03, 201.19it/s]
 11%|█▏        | 83/722 [00:00<00:03, 201.47it/s]
 14%|█▍        | 104/722 [00:00<00:03, 201.80it/s]
 17%|█▋        | 125/722 [00:00<00:02, 202.00it/s]
 20%|██        | 146/722 [00:00<00:02, 202.34it/s]
 23%|██▎       | 167/722 [00:00<00:02, 200.82it/s]
 26%|██▌       | 188/722 [00:00<00:02, 201.69it/s]
 29%|██▉       | 209/722 [00:01<00:02, 202.25it/s]
 32%|███▏      | 230/722 [00:01<00:02, 202.86it/s]
 35%|███▍      | 251/722 [00:01<00:02, 202.38it/s]
 38%|███▊      | 272/722 [00:01<00:02, 202.97it/s]
 41%|████      | 293/722 [00:01<00:02, 202.70it/s]
 43%|████▎     | 314/722 [00:01<00:02, 203.24it/s]
 46%|████▋     | 335/722 [00:01<00:01, 203.84it/s]
 49%|████▉     | 356/722 [00:01<00:01, 202.98it/s]
 52%|█████▏    | 377/722 [00:01<00:01, 203.73it/s]
 55%|█████▌    | 398/722 [00:01<00:01, 204.20it/s]
 58%|█████▊    | 419/722 [00:02<00:01, 204.57it/s]
 61%|██████    | 440/722 [00:02<00:01, 204.92it/s]
 64%|██████▍   | 461/722 [00:02<00:01, 205.00it/s]
 67%|██████▋   | 482/722 [00:02<00:01, 205.09it/s]
 70%|██████▉   | 503/722 [00:02<00:01, 205.23it/s]
 73%|███████▎  | 524/722 [00:02<00:00, 205.47it/s]
 75%|███████▌  | 545/722 [00:02<00:00, 205.81it/s]
 78%|███████▊  | 566/722 [00:02<00:00, 205.51it/s]
 81%|████████▏ | 587/722 [00:02<00:00, 205.69it/s]
 84%|████████▍ | 608/722 [00:02<00:00, 205.87it/s]
 87%|████████▋ | 629/722 [00:03<00:00, 205.84it/s]
 90%|█████████ | 650/722 [00:03<00:00, 205.98it/s]
 93%|█████████▎| 671/722 [00:03<00:00, 205.81it/s]
 96%|█████████▌| 692/722 [00:03<00:00, 205.95it/s]
 99%|█████████▉| 713/722 [00:03<00:00, 206.15it/s]
+I0221 20:32:36.015754 11685 evaluator.py:431] Running generate_until requests
+[20:32:36]   generate_until: 8666 requests, max_gen_toks=64
+Terminated
+[2026-02-21 20:38:31] Generative tasks done!
+[2026-02-21 20:38:31] ============================================
+[2026-02-21 20:38:31]   GENERATIVE RESULTS
+[2026-02-21 20:38:31] ============================================
+Traceback (most recent call last):
+  File "/dev/shm/eval/print_results.py", line 6, in <module>
+    with open(path) as f:
+         ^^^^^^^^^^
+FileNotFoundError: [Errno 2] No such file or directory: '/dev/shm/eval/results_gen/full_results.json'
+[2026-02-21 20:38:31] Gen results not found
+[2026-02-21 20:38:31] ============================================
+[2026-02-21 20:38:31]   STARTING: Remaining tasks (13)
+[2026-02-21 20:38:31] ============================================
+I0221 20:38:33.193227 12267 utils.py:148] Note: detected 192 virtual cores but NumExpr set to maximum of 64, check "NUMEXPR_MAX_THREADS" environment variable.
+I0221 20:38:33.193303 12267 utils.py:151] Note: NumExpr detected 192 cores but "NUMEXPR_MAX_THREADS" not set, so enforcing safe limit of 16.
+I0221 20:38:33.193339 12267 utils.py:164] NumExpr defaulting to 16 threads.
+I0221 20:38:33.274195 12267 config.py:58] PyTorch version 2.10.0+cu126 available.
+W0221 20:38:33.399335 12267 warnings.py:112] /dev/shm/eval/quip-sharp/lib/codebook/__init__.py:6: FutureWarning: `torch.library.impl_abstract` was renamed to `torch.library.register_fake`. Please use that instead; we will remove `torch.library.impl_abstract` in a future version of PyTorch.
+  @torch.library.impl_abstract("quip_lib::decode_matvec_e8p")
+
+W0221 20:38:33.435994 12267 warnings.py:112] /dev/shm/eval/quip-sharp/lib/codebook/__init__.py:25: FutureWarning: `torch.library.impl_abstract` was renamed to `torch.library.register_fake`. Please use that instead; we will remove `torch.library.impl_abstract` in a future version of PyTorch.
+  @torch.library.impl_abstract("quip_lib::decompress_packed_e8p")
+
+W0221 20:38:33.593651 12267 warnings.py:112] /dev/shm/eval/quip-sharp/lib/utils/matmul_had.py:96: FutureWarning: `torch.library.impl_abstract` was renamed to `torch.library.register_fake`. Please use that instead; we will remove `torch.library.impl_abstract` in a future version of PyTorch.
+  @torch.library.impl_abstract("quip_lib::hadamard")
+
+Terminated
+[2026-02-21 20:38:39] Remaining tasks done!
+[2026-02-21 20:38:39] ============================================
+[2026-02-21 20:38:39]   REMAINING RESULTS
+[2026-02-21 20:38:39] ============================================
+Traceback (most recent call last):
+  File "/dev/shm/eval/print_results.py", line 6, in <module>
+    with open(path) as f:
+         ^^^^^^^^^^
+FileNotFoundError: [Errno 2] No such file or directory: '/dev/shm/eval/results_remaining/full_results.json'
+[2026-02-21 20:38:39] Remaining results not found
+[2026-02-21 20:38:39] ============================================
+[2026-02-21 20:38:39]   ALL COMPLETE
+[2026-02-21 20:38:39] ============================================
+[2026-02-21 20:38:39] MC:        /dev/shm/eval/results_mc/full_results.json
+[2026-02-21 20:38:39] Generative:/dev/shm/eval/results_gen/full_results.json
+[2026-02-21 20:38:39] Remaining: /dev/shm/eval/results_remaining/full_results.json

variant_a/logs/eval_full_mc.log

@@ -0,0 +1,305 @@
+=== RUN 1: MC tasks (loglikelihood only) ===
+I0221 16:04:57.513942 9529 utils.py:148] Note: detected 192 virtual cores but NumExpr set to maximum of 64, check "NUMEXPR_MAX_THREADS" environment variable.
+I0221 16:04:57.514031 9529 utils.py:151] Note: NumExpr detected 192 cores but "NUMEXPR_MAX_THREADS" not set, so enforcing safe limit of 16.
+I0221 16:04:57.514069 9529 utils.py:164] NumExpr defaulting to 16 threads.
+I0221 16:04:57.595988 9529 config.py:58] PyTorch version 2.10.0+cu126 available.
+W0221 16:04:57.723166 9529 warnings.py:112] /dev/shm/eval/quip-sharp/lib/codebook/__init__.py:6: FutureWarning: `torch.library.impl_abstract` was renamed to `torch.library.register_fake`. Please use that instead; we will remove `torch.library.impl_abstract` in a future version of PyTorch.
+  @torch.library.impl_abstract("quip_lib::decode_matvec_e8p")
+
+W0221 16:04:57.761159 9529 warnings.py:112] /dev/shm/eval/quip-sharp/lib/codebook/__init__.py:25: FutureWarning: `torch.library.impl_abstract` was renamed to `torch.library.register_fake`. Please use that instead; we will remove `torch.library.impl_abstract` in a future version of PyTorch.
+  @torch.library.impl_abstract("quip_lib::decompress_packed_e8p")
+
+W0221 16:04:57.926314 9529 warnings.py:112] /dev/shm/eval/quip-sharp/lib/utils/matmul_had.py:96: FutureWarning: `torch.library.impl_abstract` was renamed to `torch.library.register_fake`. Please use that instead; we will remove `torch.library.impl_abstract` in a future version of PyTorch.
+  @torch.library.impl_abstract("quip_lib::hadamard")
+
+W0221 16:05:22.159626 9529 warnings.py:112] /dev/shm/eval/lm-evaluation-harness/lm_eval/filters/extraction.py:98: SyntaxWarning: invalid escape sequence '\s'
+  - step 2 : We parse the choice with regex :[\s]*([A-?]), where ? varies by number of choices.
+
+W0221 16:05:22.159894 9529 warnings.py:112] /dev/shm/eval/lm-evaluation-harness/lm_eval/filters/extraction.py:168: SyntaxWarning: invalid escape sequence '\s'
+  f":[\s]*({without_paren_fallback_regex})"
+
+[16:05:22] ============================================================
+[16:05:22]   Polish LLM Leaderboard Eval
+[16:05:22]   Model: QuIP# Bielik-Q2-Sharp Variant A
+[16:05:22]   lm-eval API: new
+[16:05:22]   Tasks: ['polemo2_in_multiple_choice', 'polemo2_out_multiple_choice', 'polish_8tags_multiple_choice', 'polish_cbd_multiple_choice', 'polish_ppc_multiple_choice', 'polish_psc_multiple_choice']
+[16:05:22]   Few-shot: 5
+[16:05:22] ============================================================
+[16:05:22] Loading QuIP# model from /dev/shm/eval/model...
+I0221 16:05:24.190070 9529 modeling.py:987] We will use 90% of the memory on device 0 for storing the model, and 10% for the buffer to avoid OOM. You can set `max_memory` in to a higher value to use more memory (at your own risk).
+[16:05:25] Model loaded in 3.3s
+[16:05:25] Tokenizer: speakleash/Bielik-11B-v2.3-Instruct (vocab=32000)
+[16:05:25] 
+Running 6 tasks...
+I0221 16:05:25.896242 9529 evaluator.py:152] Setting random seed to 0 | Setting numpy seed to 1234 | Setting torch manual seed to 1234
+I0221 16:05:25.896321 9529 evaluator.py:203] Using pre-initialized model
+W0221 16:05:43.775950 9529 evaluator.py:251] Overwriting default num_fewshot of polish_psc_multiple_choice from None to 5
+I0221 16:05:43.776097 9529 evaluator.py:261] Setting fewshot random generator seed to 1234
+W0221 16:05:43.776127 9529 evaluator.py:251] Overwriting default num_fewshot of polish_ppc_multiple_choice from None to 5
+I0221 16:05:43.776158 9529 evaluator.py:261] Setting fewshot random generator seed to 1234
+W0221 16:05:43.776176 9529 evaluator.py:251] Overwriting default num_fewshot of polish_cbd_multiple_choice from None to 5
+I0221 16:05:43.776201 9529 evaluator.py:261] Setting fewshot random generator seed to 1234
+W0221 16:05:43.776215 9529 evaluator.py:251] Overwriting default num_fewshot of polish_8tags_multiple_choice from None to 5
+I0221 16:05:43.776236 9529 evaluator.py:261] Setting fewshot random generator seed to 1234
+W0221 16:05:43.776249 9529 evaluator.py:251] Overwriting default num_fewshot of polemo2_out_multiple_choice from None to 5
+I0221 16:05:43.776278 9529 evaluator.py:261] Setting fewshot random generator seed to 1234
+W0221 16:05:43.776291 9529 evaluator.py:251] Overwriting default num_fewshot of polemo2_in_multiple_choice from None to 5
+I0221 16:05:43.776311 9529 evaluator.py:261] Setting fewshot random generator seed to 1234
+I0221 16:05:43.777427 9529 task.py:410] Building contexts for polish_psc_multiple_choice on rank 0...
+
  0%|          | 0/1078 [00:00<?, ?it/s]
  2%|▏         | 24/1078 [00:00<00:04, 237.74it/s]
  5%|▍         | 50/1078 [00:00<00:04, 246.89it/s]
  7%|▋         | 76/1078 [00:00<00:04, 249.65it/s]
  9%|▉         | 102/1078 [00:00<00:03, 251.04it/s]
 12%|█▏        | 128/1078 [00:00<00:03, 251.65it/s]
 14%|█▍        | 154/1078 [00:00<00:03, 252.02it/s]
 17%|█▋        | 180/1078 [00:00<00:03, 252.11it/s]
 19%|█▉        | 206/1078 [00:00<00:03, 252.48it/s]
 22%|██▏       | 232/1078 [00:00<00:03, 252.69it/s]
 24%|██▍       | 258/1078 [00:01<00:03, 252.99it/s]
 26%|██▋       | 284/1078 [00:01<00:03, 252.74it/s]
 29%|██▉       | 310/1078 [00:01<00:03, 242.45it/s]
 31%|███       | 335/1078 [00:01<00:03, 239.17it/s]
 33%|███▎      | 361/1078 [00:01<00:02, 241.82it/s]
 36%|███▌      | 386/1078 [00:01<00:02, 242.03it/s]
 38%|███▊      | 411/1078 [00:01<00:02, 233.63it/s]
 40%|████      | 435/1078 [00:01<00:02, 234.02it/s]
 43%|████▎     | 459/1078 [00:01<00:02, 233.70it/s]
 45%|████▍     | 483/1078 [00:01<00:02, 231.61it/s]
 47%|████▋     | 509/1078 [00:02<00:02, 237.98it/s]
 50%|████▉     | 535/1078 [00:02<00:02, 242.43it/s]
 52%|█████▏    | 561/1078 [00:02<00:02, 245.68it/s]
 54%|█████▍    | 587/1078 [00:02<00:01, 247.97it/s]
 57%|█████▋    | 613/1078 [00:02<00:01, 249.43it/s]
 59%|█████▉    | 639/1078 [00:02<00:01, 250.61it/s]
 62%|██████▏   | 665/1078 [00:02<00:01, 251.45it/s]
 64%|██████▍   | 691/1078 [00:02<00:01, 252.04it/s]
 67%|██████▋   | 717/1078 [00:02<00:01, 252.17it/s]
 69%|██████▉   | 743/1078 [00:03<00:01, 252.42it/s]
 71%|███████▏  | 769/1078 [00:03<00:01, 253.04it/s]
 74%|███████▎  | 795/1078 [00:03<00:01, 253.16it/s]
 76%|███████▌  | 821/1078 [00:03<00:01, 249.00it/s]
 79%|███████▊  | 847/1078 [00:03<00:00, 249.29it/s]
 81%|████████  | 873/1078 [00:03<00:00, 250.33it/s]
 83%|████████▎ | 899/1078 [00:03<00:00, 250.39it/s]
 86%|████████▌ | 925/1078 [00:03<00:00, 250.52it/s]
 88%|████████▊ | 951/1078 [00:03<00:00, 250.56it/s]
 91%|█████████ | 977/1078 [00:03<00:00, 250.73it/s]
 93%|█████████▎| 1003/1078 [00:04<00:00, 250.92it/s]
 95%|█████████▌| 1029/1078 [00:04<00:00, 251.03it/s]
 98%|█████████▊| 1055/1078 [00:04<00:00, 251.42it/s]
+I0221 16:05:48.163607 9529 task.py:410] Building contexts for polish_ppc_multiple_choice on rank 0...
+
  0%|          | 0/1000 [00:00<?, ?it/s]
  2%|▏         | 21/1000 [00:00<00:04, 207.98it/s]
  4%|▍         | 44/1000 [00:00<00:04, 218.86it/s]
  7%|▋         | 67/1000 [00:00<00:04, 219.32it/s]
  9%|▉         | 89/1000 [00:00<00:04, 212.61it/s]
 11%|█         | 111/1000 [00:00<00:04, 208.12it/s]
 13%|█▎        | 132/1000 [00:00<00:04, 205.97it/s]
 15%|█▌        | 153/1000 [00:00<00:04, 206.85it/s]
 18%|█▊        | 176/1000 [00:00<00:03, 213.05it/s]
 20%|█▉        | 199/1000 [00:00<00:03, 215.05it/s]
 22%|██▏       | 221/1000 [00:01<00:03, 211.78it/s]
 24%|██▍       | 243/1000 [00:01<00:03, 210.87it/s]
 26%|██▋       | 265/1000 [00:01<00:03, 210.63it/s]
 29%|██▊       | 287/1000 [00:01<00:03, 210.81it/s]
 31%|███       | 310/1000 [00:01<00:03, 215.21it/s]
 33%|███▎      | 333/1000 [00:01<00:03, 219.44it/s]
 36%|███▌      | 356/1000 [00:01<00:02, 222.19it/s]
 38%|███▊      | 379/1000 [00:01<00:02, 224.28it/s]
 40%|████      | 402/1000 [00:01<00:02, 225.79it/s]
 42%|████▎     | 425/1000 [00:01<00:02, 226.99it/s]
 45%|████▍     | 448/1000 [00:02<00:02, 227.70it/s]
 47%|████▋     | 471/1000 [00:02<00:02, 225.16it/s]
 49%|████▉     | 494/1000 [00:02<00:02, 226.39it/s]
 52%|█████▏    | 517/1000 [00:02<00:02, 227.34it/s]
 54%|█████▍    | 540/1000 [00:02<00:02, 227.79it/s]
 56%|█████▋    | 563/1000 [00:02<00:01, 228.24it/s]
 59%|█████▊    | 586/1000 [00:02<00:01, 228.53it/s]
 61%|██████    | 609/1000 [00:02<00:01, 228.73it/s]
 63%|██████▎   | 632/1000 [00:02<00:01, 221.86it/s]
 66%|██████▌   | 655/1000 [00:02<00:01, 224.12it/s]
 68%|██████▊   | 678/1000 [00:03<00:01, 225.78it/s]
 70%|███████   | 701/1000 [00:03<00:01, 226.61it/s]
 72%|███████▏  | 724/1000 [00:03<00:01, 227.47it/s]
 75%|███████▍  | 747/1000 [00:03<00:01, 226.87it/s]
 77%|███████▋  | 770/1000 [00:03<00:01, 224.99it/s]
 79%|███████▉  | 793/1000 [00:03<00:00, 220.55it/s]
 82%|████████▏ | 816/1000 [00:03<00:00, 216.20it/s]
 84%|████████▍ | 839/1000 [00:03<00:00, 219.14it/s]
 86%|████████▌ | 861/1000 [00:03<00:00, 214.99it/s]
 88%|████████▊ | 883/1000 [00:04<00:00, 211.42it/s]
 91%|█████████ | 906/1000 [00:04<00:00, 216.59it/s]
 93%|█████████▎| 928/1000 [00:04<00:00, 207.62it/s]
 95%|█████████▌| 951/1000 [00:04<00:00, 213.36it/s]
 97%|█████████▋| 974/1000 [00:04<00:00, 218.08it/s]
+I0221 16:05:52.734501 9529 task.py:410] Building contexts for polish_cbd_multiple_choice on rank 0...
+
  0%|          | 0/1000 [00:00<?, ?it/s]
  2%|▏         | 20/1000 [00:00<00:05, 191.33it/s]
  4%|▍         | 40/1000 [00:00<00:05, 191.26it/s]
  6%|▌         | 60/1000 [00:00<00:04, 191.56it/s]
  8%|▊         | 80/1000 [00:00<00:04, 191.34it/s]
 10%|█         | 100/1000 [00:00<00:04, 191.41it/s]
 12%|█▏        | 120/1000 [00:00<00:04, 191.43it/s]
 14%|█▍        | 140/1000 [00:00<00:04, 188.74it/s]
 16%|█▌        | 160/1000 [00:00<00:04, 189.77it/s]
 18%|█▊        | 180/1000 [00:00<00:04, 190.45it/s]
 20%|██        | 200/1000 [00:01<00:04, 190.33it/s]
 22%|██▏       | 220/1000 [00:01<00:04, 190.36it/s]
 24%|██▍       | 240/1000 [00:01<00:03, 190.44it/s]
 26%|██▌       | 260/1000 [00:01<00:03, 190.57it/s]
 28%|██▊       | 280/1000 [00:01<00:03, 190.68it/s]
 30%|███       | 300/1000 [00:01<00:03, 190.77it/s]
 32%|███▏      | 320/1000 [00:01<00:03, 190.86it/s]
 34%|███▍      | 340/1000 [00:01<00:03, 190.95it/s]
 36%|███▌      | 360/1000 [00:01<00:03, 190.78it/s]
 38%|███▊      | 380/1000 [00:01<00:03, 190.95it/s]
 40%|████      | 400/1000 [00:02<00:03, 191.15it/s]
 42%|████▏     | 420/1000 [00:02<00:03, 191.34it/s]
 44%|████▍     | 440/1000 [00:02<00:02, 188.48it/s]
 46%|████▌     | 460/1000 [00:02<00:02, 189.01it/s]
 48%|████▊     | 480/1000 [00:02<00:02, 189.57it/s]
 50%|█████     | 500/1000 [00:02<00:02, 190.03it/s]
 52%|█████▏    | 520/1000 [00:02<00:02, 190.47it/s]
 54%|█████▍    | 540/1000 [00:02<00:02, 190.48it/s]
 56%|█████▌    | 560/1000 [00:02<00:02, 190.38it/s]
 58%|█████▊    | 580/1000 [00:03<00:02, 189.99it/s]
 60%|█████▉    | 599/1000 [00:03<00:02, 189.83it/s]
 62%|██████▏   | 618/1000 [00:03<00:02, 189.64it/s]
 64%|██████▍   | 638/1000 [00:03<00:01, 189.95it/s]
 66%|██████▌   | 657/1000 [00:03<00:01, 189.88it/s]
 68%|██████▊   | 676/1000 [00:03<00:01, 189.84it/s]
 70%|██████▉   | 695/1000 [00:03<00:01, 189.64it/s]
 72%|███████▏  | 715/1000 [00:03<00:01, 189.94it/s]
 74%|███████▎  | 735/1000 [00:03<00:01, 190.09it/s]
 76%|███████▌  | 755/1000 [00:03<00:01, 189.75it/s]
 78%|███████▊  | 775/1000 [00:04<00:01, 190.04it/s]
 80%|███████▉  | 795/1000 [00:04<00:01, 189.91it/s]
 82%|████████▏ | 815/1000 [00:04<00:00, 190.06it/s]
 84%|████████▎ | 835/1000 [00:04<00:00, 189.75it/s]
 86%|████████▌ | 855/1000 [00:04<00:00, 189.97it/s]
 88%|████████▊ | 875/1000 [00:04<00:00, 190.12it/s]
 90%|████████▉ | 895/1000 [00:04<00:00, 190.22it/s]
 92%|█████████▏| 915/1000 [00:04<00:00, 190.62it/s]
 94%|█████████▎| 935/1000 [00:04<00:00, 190.67it/s]
 96%|█████████▌| 955/1000 [00:05<00:00, 190.38it/s]
 98%|█████████▊| 975/1000 [00:05<00:00, 189.98it/s]
 99%|█████████▉| 994/1000 [00:05<00:00, 189.82it/s]
+I0221 16:05:58.015220 9529 task.py:410] Building contexts for polish_8tags_multiple_choice on rank 0...
+
  0%|          | 0/4372 [00:00<?, ?it/s]
  1%|          | 28/4372 [00:00<00:15, 273.81it/s]
  1%|▏         | 57/4372 [00:00<00:15, 277.37it/s]
  2%|▏         | 86/4372 [00:00<00:15, 279.03it/s]
  3%|▎         | 115/4372 [00:00<00:15, 280.16it/s]
  3%|▎         | 144/4372 [00:00<00:15, 280.58it/s]
  4%|▍         | 173/4372 [00:00<00:14, 281.10it/s]
  5%|▍         | 202/4372 [00:00<00:14, 282.00it/s]
  5%|▌         | 231/4372 [00:00<00:14, 282.41it/s]
  6%|▌         | 260/4372 [00:00<00:14, 282.70it/s]
  7%|▋         | 289/4372 [00:01<00:14, 282.64it/s]
  7%|▋         | 318/4372 [00:01<00:14, 282.49it/s]
  8%|▊         | 347/4372 [00:01<00:14, 282.15it/s]
  9%|▊         | 376/4372 [00:01<00:14, 282.46it/s]
  9%|▉         | 405/4372 [00:01<00:14, 282.70it/s]
 10%|▉         | 434/4372 [00:01<00:13, 282.57it/s]
 11%|█         | 463/4372 [00:01<00:13, 282.95it/s]
 11%|█▏        | 492/4372 [00:01<00:13, 283.16it/s]
 12%|█▏        | 521/4372 [00:01<00:13, 282.91it/s]
 13%|█▎        | 550/4372 [00:01<00:13, 283.54it/s]
 13%|█▎        | 579/4372 [00:02<00:13, 282.96it/s]
 14%|█▍        | 608/4372 [00:02<00:13, 282.94it/s]
 15%|█▍        | 637/4372 [00:02<00:13, 282.19it/s]
 15%|█▌        | 666/4372 [00:02<00:13, 282.27it/s]
 16%|█▌        | 695/4372 [00:02<00:13, 282.07it/s]
 17%|█▋        | 724/4372 [00:02<00:12, 282.15it/s]
 17%|█▋        | 753/4372 [00:02<00:12, 282.06it/s]
 18%|█▊        | 782/4372 [00:02<00:12, 280.48it/s]
 19%|█▊        | 811/4372 [00:02<00:12, 280.82it/s]
 19%|█▉        | 840/4372 [00:02<00:12, 281.45it/s]
 20%|█▉        | 869/4372 [00:03<00:12, 281.61it/s]
 21%|██        | 898/4372 [00:03<00:12, 281.59it/s]
 21%|██        | 927/4372 [00:03<00:12, 281.20it/s]
 22%|██▏       | 956/4372 [00:03<00:12, 281.06it/s]
 23%|██▎       | 985/4372 [00:03<00:12, 281.02it/s]
 23%|██▎       | 1014/4372 [00:03<00:11, 281.32it/s]
 24%|██▍       | 1043/4372 [00:03<00:11, 282.12it/s]
 25%|██▍       | 1072/4372 [00:03<00:11, 282.39it/s]
 25%|██▌       | 1101/4372 [00:03<00:11, 282.47it/s]
 26%|██▌       | 1130/4372 [00:04<00:11, 282.74it/s]
 27%|██▋       | 1159/4372 [00:04<00:11, 283.08it/s]
 27%|██▋       | 1188/4372 [00:04<00:11, 283.08it/s]
 28%|██▊       | 1217/4372 [00:04<00:11, 283.41it/s]
 28%|██▊       | 1246/4372 [00:04<00:11, 282.36it/s]
 29%|██▉       | 1275/4372 [00:04<00:10, 281.81it/s]
 30%|██▉       | 1304/4372 [00:04<00:10, 281.93it/s]
 30%|███       | 1333/4372 [00:04<00:10, 282.48it/s]
 31%|███       | 1362/4372 [00:04<00:10, 283.11it/s]
 32%|███▏      | 1391/4372 [00:04<00:10, 283.27it/s]
 32%|███▏      | 1420/4372 [00:05<00:10, 283.53it/s]
 33%|███▎      | 1449/4372 [00:05<00:10, 283.77it/s]
 34%|███▍      | 1478/4372 [00:05<00:10, 283.89it/s]
 34%|███▍      | 1507/4372 [00:05<00:10, 283.72it/s]
 35%|███▌      | 1536/4372 [00:05<00:09, 283.89it/s]
 36%|███▌      | 1565/4372 [00:05<00:09, 283.71it/s]
 36%|███▋      | 1594/4372 [00:05<00:09, 283.81it/s]
 37%|███▋      | 1623/4372 [00:05<00:09, 283.96it/s]
 38%|███▊      | 1652/4372 [00:05<00:09, 283.15it/s]
 38%|███▊      | 1681/4372 [00:06<00:16, 165.12it/s]
 39%|███▉      | 1709/4372 [00:06<00:14, 187.40it/s]
 40%|███▉      | 1738/4372 [00:06<00:12, 208.32it/s]
 40%|████      | 1767/4372 [00:06<00:11, 225.87it/s]
 41%|████      | 1796/4372 [00:06<00:10, 240.12it/s]
 42%|████▏     | 1825/4372 [00:06<00:10, 251.32it/s]
 42%|████▏     | 1854/4372 [00:06<00:09, 259.54it/s]
 43%|████▎     | 1883/4372 [00:06<00:09, 265.96it/s]
 44%|████▎     | 1912/4372 [00:07<00:09, 270.68it/s]
 44%|████▍     | 1941/4372 [00:07<00:08, 274.23it/s]
 45%|████▌     | 1970/4372 [00:07<00:08, 276.44it/s]
 46%|████▌     | 1999/4372 [00:07<00:08, 278.11it/s]
 46%|████▋     | 2028/4372 [00:07<00:08, 279.15it/s]
 47%|████▋     | 2057/4372 [00:07<00:08, 279.93it/s]
 48%|████▊     | 2086/4372 [00:07<00:08, 280.85it/s]
 48%|████▊     | 2115/4372 [00:07<00:08, 281.57it/s]
 49%|████▉     | 2144/4372 [00:07<00:07, 281.72it/s]
 50%|████▉     | 2173/4372 [00:07<00:07, 281.93it/s]
 50%|█████     | 2202/4372 [00:08<00:07, 281.90it/s]
 51%|█████     | 2231/4372 [00:08<00:07, 281.92it/s]
 52%|█████▏    | 2260/4372 [00:08<00:07, 282.05it/s]
 52%|█████▏    | 2289/4372 [00:08<00:07, 282.49it/s]
 53%|█████▎    | 2318/4372 [00:08<00:07, 282.23it/s]
 54%|█████▎    | 2347/4372 [00:08<00:07, 282.34it/s]
 54%|█████▍    | 2376/4372 [00:08<00:07, 282.40it/s]
 55%|█████▌    | 2405/4372 [00:08<00:06, 282.38it/s]
 56%|█████▌    | 2434/4372 [00:08<00:06, 281.78it/s]
 56%|█████▋    | 2463/4372 [00:08<00:06, 279.11it/s]
 57%|█████▋    | 2492/4372 [00:09<00:06, 279.34it/s]
 58%|█████▊    | 2521/4372 [00:09<00:06, 279.94it/s]
 58%|█████▊    | 2550/4372 [00:09<00:06, 280.61it/s]
 59%|█████▉    | 2579/4372 [00:09<00:06, 280.80it/s]
 60%|█████▉    | 2608/4372 [00:09<00:06, 280.48it/s]
 60%|██████    | 2637/4372 [00:09<00:06, 280.77it/s]
 61%|██████    | 2666/4372 [00:09<00:06, 272.35it/s]
 62%|██████▏   | 2694/4372 [00:09<00:06, 270.38it/s]
 62%|██████▏   | 2722/4372 [00:09<00:06, 269.03it/s]
 63%|██████▎   | 2749/4372 [00:10<00:06, 268.78it/s]
 64%|██████▎   | 2778/4372 [00:10<00:05, 273.08it/s]
 64%|██████▍   | 2807/4372 [00:10<00:05, 276.14it/s]
 65%|██████▍   | 2836/4372 [00:10<00:05, 277.89it/s]
 66%|██████▌   | 2864/4372 [00:10<00:05, 277.87it/s]
 66%|██████▌   | 2893/4372 [00:10<00:05, 279.38it/s]
 67%|██████▋   | 2922/4372 [00:10<00:05, 280.57it/s]
 67%|██████▋   | 2951/4372 [00:10<00:05, 281.49it/s]
 68%|██████▊   | 2980/4372 [00:10<00:04, 282.17it/s]
 69%|██████▉   | 3009/4372 [00:10<00:04, 282.36it/s]
 69%|██████▉   | 3038/4372 [00:11<00:04, 281.42it/s]
 70%|███████   | 3067/4372 [00:11<00:04, 281.24it/s]
 71%|███████   | 3096/4372 [00:11<00:04, 281.04it/s]
 71%|███████▏  | 3125/4372 [00:11<00:04, 281.54it/s]
 72%|███████▏  | 3154/4372 [00:11<00:04, 281.39it/s]
 73%|███████▎  | 3183/4372 [00:11<00:04, 281.65it/s]
 73%|███████▎  | 3212/4372 [00:11<00:04, 281.79it/s]
 74%|███████▍  | 3241/4372 [00:11<00:04, 281.74it/s]
 75%|███████▍  | 3270/4372 [00:11<00:03, 281.53it/s]
 75%|███████▌  | 3299/4372 [00:11<00:03, 280.77it/s]
 76%|███████▌  | 3328/4372 [00:12<00:03, 280.15it/s]
 77%|███████▋  | 3357/4372 [00:12<00:03, 278.66it/s]
 77%|███████▋  | 3386/4372 [00:12<00:03, 279.48it/s]
 78%|███████▊  | 3415/4372 [00:12<00:03, 280.28it/s]
 79%|███████▉  | 3444/4372 [00:12<00:03, 280.86it/s]
 79%|███████▉  | 3473/4372 [00:12<00:03, 281.16it/s]
 80%|████████  | 3502/4372 [00:12<00:03, 281.54it/s]
 81%|████████  | 3531/4372 [00:12<00:02, 282.34it/s]
 81%|████████▏ | 3560/4372 [00:12<00:02, 282.60it/s]
 82%|████████▏ | 3589/4372 [00:13<00:02, 282.55it/s]
 83%|████████▎ | 3618/4372 [00:13<00:02, 281.80it/s]
 83%|████████▎ | 3647/4372 [00:13<00:02, 281.81it/s]
 84%|████████▍ | 3676/4372 [00:13<00:02, 281.82it/s]
 85%|████████▍ | 3705/4372 [00:13<00:02, 282.54it/s]
 85%|████████▌ | 3734/4372 [00:13<00:02, 282.76it/s]
 86%|████████▌ | 3763/4372 [00:13<00:02, 282.98it/s]
 87%|████████▋ | 3792/4372 [00:13<00:02, 283.06it/s]
 87%|████████▋ | 3821/4372 [00:13<00:01, 283.07it/s]
 88%|████████▊ | 3850/4372 [00:13<00:01, 282.92it/s]
 89%|████████▊ | 3879/4372 [00:14<00:01, 282.83it/s]
 89%|████████▉ | 3908/4372 [00:14<00:01, 282.72it/s]
 90%|█████████ | 3937/4372 [00:14<00:01, 282.56it/s]
 91%|█████████ | 3966/4372 [00:14<00:01, 282.92it/s]
 91%|█████████▏| 3995/4372 [00:14<00:01, 282.47it/s]
 92%|█████████▏| 4024/4372 [00:14<00:01, 282.68it/s]
 93%|█████████▎| 4053/4372 [00:14<00:01, 282.57it/s]
 93%|█████████▎| 4082/4372 [00:14<00:01, 282.58it/s]
 94%|█████████▍| 4111/4372 [00:14<00:00, 282.56it/s]
 95%|█████████▍| 4140/4372 [00:14<00:00, 281.75it/s]
 95%|█████████▌| 4169/4372 [00:15<00:00, 281.97it/s]
 96%|█████████▌| 4198/4372 [00:15<00:00, 281.93it/s]
 97%|█████████▋| 4227/4372 [00:15<00:00, 281.94it/s]
 97%|█████████▋| 4256/4372 [00:15<00:00, 281.79it/s]
 98%|█████████▊| 4285/4372 [00:15<00:00, 281.88it/s]
 99%|█████████▊| 4314/4372 [00:15<00:00, 281.95it/s]
 99%|█████████▉| 4343/4372 [00:15<00:00, 281.77it/s]
+I0221 16:06:13.846424 9529 task.py:410] Building contexts for polemo2_out_multiple_choice on rank 0...
+
  0%|          | 0/494 [00:00<?, ?it/s]
  5%|▍         | 24/494 [00:00<00:02, 234.99it/s]
 10%|▉         | 48/494 [00:00<00:01, 234.97it/s]
 15%|█▍        | 72/494 [00:00<00:01, 235.48it/s]
 19%|█▉        | 96/494 [00:00<00:01, 235.61it/s]
 24%|██▍       | 120/494 [00:00<00:01, 235.85it/s]
 29%|██▉       | 144/494 [00:00<00:01, 235.95it/s]
 34%|███▍      | 168/494 [00:00<00:01, 235.56it/s]
 39%|███▉      | 192/494 [00:00<00:01, 235.71it/s]
 44%|████▎     | 216/494 [00:00<00:01, 236.12it/s]
 49%|████▊     | 240/494 [00:01<00:01, 236.06it/s]
 53%|█████▎    | 264/494 [00:01<00:00, 236.17it/s]
 58%|█████▊    | 288/494 [00:01<00:00, 235.97it/s]
 63%|██████▎   | 312/494 [00:01<00:00, 235.80it/s]
 68%|██████▊   | 336/494 [00:01<00:00, 235.39it/s]
 73%|███████▎  | 360/494 [00:01<00:00, 235.36it/s]
 78%|███████▊  | 384/494 [00:01<00:00, 235.63it/s]
 83%|████████▎ | 408/494 [00:01<00:00, 235.16it/s]
 87%|████████▋ | 432/494 [00:01<00:00, 235.08it/s]
 92%|█████████▏| 456/494 [00:01<00:00, 235.52it/s]
 97%|█████████▋| 480/494 [00:02<00:00, 235.93it/s]
+I0221 16:06:15.950167 9529 task.py:410] Building contexts for polemo2_in_multiple_choice on rank 0...
+
  0%|          | 0/722 [00:00<?, ?it/s]
  3%|▎         | 24/722 [00:00<00:02, 235.60it/s]
  7%|▋         | 48/722 [00:00<00:02, 236.32it/s]
 10%|▉         | 72/722 [00:00<00:02, 236.88it/s]
 13%|█▎        | 96/722 [00:00<00:02, 237.10it/s]
 17%|█▋        | 120/722 [00:00<00:02, 237.16it/s]
 20%|█▉        | 144/722 [00:00<00:02, 236.84it/s]
 23%|██▎       | 168/722 [00:00<00:02, 237.25it/s]
 27%|██▋       | 192/722 [00:00<00:02, 237.77it/s]
 30%|██▉       | 216/722 [00:00<00:02, 237.84it/s]
 33%|███▎      | 240/722 [00:01<00:02, 238.46it/s]
 37%|███▋      | 264/722 [00:01<00:01, 238.39it/s]
 40%|███▉      | 288/722 [00:01<00:01, 238.30it/s]
 43%|████▎     | 312/722 [00:01<00:01, 238.20it/s]
 47%|████▋     | 336/722 [00:01<00:01, 238.47it/s]
 50%|█████     | 361/722 [00:01<00:01, 239.06it/s]
 53%|█████▎    | 385/722 [00:01<00:01, 238.73it/s]
 57%|█████▋    | 409/722 [00:01<00:01, 238.83it/s]
 60%|█████▉    | 433/722 [00:01<00:01, 238.88it/s]
 63%|██████▎   | 457/722 [00:01<00:01, 238.68it/s]
 67%|██████▋   | 481/722 [00:02<00:01, 238.82it/s]
 70%|██████▉   | 505/722 [00:02<00:00, 238.87it/s]
 73%|███████▎  | 529/722 [00:02<00:00, 239.14it/s]
 77%|███████▋  | 553/722 [00:02<00:00, 239.12it/s]
 80%|███████▉  | 577/722 [00:02<00:00, 239.00it/s]
 83%|████████▎ | 601/722 [00:02<00:00, 237.63it/s]
 87%|████████▋ | 625/722 [00:02<00:00, 238.07it/s]
 90%|████████▉ | 649/722 [00:02<00:00, 238.25it/s]
 93%|█████████▎| 673/722 [00:02<00:00, 238.57it/s]
 97%|█████████▋| 697/722 [00:02<00:00, 238.80it/s]
+I0221 16:06:18.991096 9529 evaluator.py:431] Running loglikelihood requests
+[16:06:47]   loglikelihood: 50996 requests, batch_size=8 (length-sorted)
+[16:06:47]   sequence lengths: min=543, max=2048, median=659
+[16:08:13]   loglikelihood: 400/50996 (4.7 req/s, ETA 180.8min)
+[16:09:39]   loglikelihood: 800/50996 (4.7 req/s, ETA 179.4min)
+[16:11:06]   loglikelihood: 1200/50996 (4.6 req/s, ETA 178.7min)
+[16:12:32]   loglikelihood: 1600/50996 (4.6 req/s, ETA 177.4min)
+[16:14:00]   loglikelihood: 2000/50996 (4.6 req/s, ETA 176.5min)
+[16:15:27]   loglikelihood: 2400/50996 (4.6 req/s, ETA 175.4min)
+[16:16:55]   loglikelihood: 2800/50996 (4.6 req/s, ETA 174.4min)
+[16:18:23]   loglikelihood: 3200/50996 (4.6 req/s, ETA 173.3min)
+[16:19:52]   loglikelihood: 3600/50996 (4.6 req/s, ETA 172.2min)
+[16:21:20]   loglikelihood: 4000/50996 (4.6 req/s, ETA 170.9min)
+[16:22:49]   loglikelihood: 4400/50996 (4.6 req/s, ETA 169.7min)
+[16:24:18]   loglikelihood: 4800/50996 (4.6 req/s, ETA 168.6min)
+[16:25:47]   loglikelihood: 5200/50996 (4.6 req/s, ETA 167.3min)
+[16:27:17]   loglikelihood: 5600/50996 (4.6 req/s, ETA 166.1min)
+[16:28:47]   loglikelihood: 6000/50996 (4.5 req/s, ETA 164.9min)
+[16:30:17]   loglikelihood: 6400/50996 (4.5 req/s, ETA 163.7min)
+[16:31:47]   loglikelihood: 6800/50996 (4.5 req/s, ETA 162.4min)
+[16:33:17]   loglikelihood: 7200/50996 (4.5 req/s, ETA 161.2min)
+[16:34:48]   loglikelihood: 7600/50996 (4.5 req/s, ETA 159.9min)
+[16:36:19]   loglikelihood: 8000/50996 (4.5 req/s, ETA 158.7min)
+[16:37:50]   loglikelihood: 8400/50996 (4.5 req/s, ETA 157.4min)
+[16:39:21]   loglikelihood: 8800/50996 (4.5 req/s, ETA 156.1min)
+[16:40:52]   loglikelihood: 9200/50996 (4.5 req/s, ETA 154.8min)
+[16:42:24]   loglikelihood: 9600/50996 (4.5 req/s, ETA 153.5min)
+[16:43:55]   loglikelihood: 10000/50996 (4.5 req/s, ETA 152.2min)
+[16:45:27]   loglikelihood: 10400/50996 (4.5 req/s, ETA 150.9min)
+[16:46:59]   loglikelihood: 10800/50996 (4.5 req/s, ETA 149.6min)
+[16:48:31]   loglikelihood: 11200/50996 (4.5 req/s, ETA 148.3min)
+[16:50:04]   loglikelihood: 11600/50996 (4.5 req/s, ETA 147.0min)
+[16:51:36]   loglikelihood: 12000/50996 (4.5 req/s, ETA 145.6min)
+[16:53:09]   loglikelihood: 12400/50996 (4.5 req/s, ETA 144.3min)
+[16:54:41]   loglikelihood: 12800/50996 (4.5 req/s, ETA 142.9min)
+[16:56:14]   loglikelihood: 13200/50996 (4.4 req/s, ETA 141.6min)
+[16:57:47]   loglikelihood: 13600/50996 (4.4 req/s, ETA 140.2min)
+[16:59:21]   loglikelihood: 14000/50996 (4.4 req/s, ETA 138.9min)
+[17:00:54]   loglikelihood: 14400/50996 (4.4 req/s, ETA 137.5min)
+[17:02:28]   loglikelihood: 14800/50996 (4.4 req/s, ETA 136.2min)
+[17:04:01]   loglikelihood: 15200/50996 (4.4 req/s, ETA 134.8min)
+[17:05:35]   loglikelihood: 15600/50996 (4.4 req/s, ETA 133.4min)
+[17:07:09]   loglikelihood: 16000/50996 (4.4 req/s, ETA 132.0min)
+[17:08:43]   loglikelihood: 16400/50996 (4.4 req/s, ETA 130.7min)
+[17:10:18]   loglikelihood: 16800/50996 (4.4 req/s, ETA 129.3min)
+[17:11:52]   loglikelihood: 17200/50996 (4.4 req/s, ETA 127.9min)
+[17:13:27]   loglikelihood: 17600/50996 (4.4 req/s, ETA 126.5min)
+[17:15:02]   loglikelihood: 18000/50996 (4.4 req/s, ETA 125.1min)
+[17:16:37]   loglikelihood: 18400/50996 (4.4 req/s, ETA 123.7min)
+[17:18:12]   loglikelihood: 18800/50996 (4.4 req/s, ETA 122.3min)
+[17:19:47]   loglikelihood: 19200/50996 (4.4 req/s, ETA 120.9min)
+[17:21:23]   loglikelihood: 19600/50996 (4.4 req/s, ETA 119.5min)
+[17:22:58]   loglikelihood: 20000/50996 (4.4 req/s, ETA 118.1min)
+[17:24:34]   loglikelihood: 20400/50996 (4.4 req/s, ETA 116.7min)
+[17:26:10]   loglikelihood: 20800/50996 (4.4 req/s, ETA 115.2min)
+[17:27:46]   loglikelihood: 21200/50996 (4.4 req/s, ETA 113.8min)
+[17:29:22]   loglikelihood: 21600/50996 (4.4 req/s, ETA 112.4min)
+[17:30:59]   loglikelihood: 22000/50996 (4.4 req/s, ETA 111.0min)
+[17:32:35]   loglikelihood: 22400/50996 (4.4 req/s, ETA 109.5min)
+[17:34:12]   loglikelihood: 22800/50996 (4.3 req/s, ETA 108.1min)
+[17:35:49]   loglikelihood: 23200/50996 (4.3 req/s, ETA 106.7min)
+[17:37:26]   loglikelihood: 23600/50996 (4.3 req/s, ETA 105.2min)
+[17:39:04]   loglikelihood: 24000/50996 (4.3 req/s, ETA 103.8min)
+[17:40:41]   loglikelihood: 24400/50996 (4.3 req/s, ETA 102.4min)
+[17:42:19]   loglikelihood: 24800/50996 (4.3 req/s, ETA 100.9min)
+[17:43:57]   loglikelihood: 25200/50996 (4.3 req/s, ETA 99.5min)
+[17:45:36]   loglikelihood: 25600/50996 (4.3 req/s, ETA 98.0min)
+[17:47:14]   loglikelihood: 26000/50996 (4.3 req/s, ETA 96.6min)
+[17:48:53]   loglikelihood: 26400/50996 (4.3 req/s, ETA 95.1min)
+[17:50:32]   loglikelihood: 26800/50996 (4.3 req/s, ETA 93.7min)
+[17:52:11]   loglikelihood: 27200/50996 (4.3 req/s, ETA 92.2min)
+[17:53:50]   loglikelihood: 27600/50996 (4.3 req/s, ETA 90.7min)
+[17:55:30]   loglikelihood: 28000/50996 (4.3 req/s, ETA 89.3min)
+[17:57:10]   loglikelihood: 28400/50996 (4.3 req/s, ETA 87.8min)
+[17:58:49]   loglikelihood: 28800/50996 (4.3 req/s, ETA 86.3min)
+[18:00:29]   loglikelihood: 29200/50996 (4.3 req/s, ETA 84.9min)
+[18:02:10]   loglikelihood: 29600/50996 (4.3 req/s, ETA 83.4min)
+[18:03:50]   loglikelihood: 30000/50996 (4.3 req/s, ETA 81.9min)
+[18:05:31]   loglikelihood: 30400/50996 (4.3 req/s, ETA 80.4min)
+[18:07:13]   loglikelihood: 30800/50996 (4.3 req/s, ETA 79.0min)
+[18:08:54]   loglikelihood: 31200/50996 (4.3 req/s, ETA 77.5min)
+[18:10:36]   loglikelihood: 31600/50996 (4.3 req/s, ETA 76.0min)
+[18:12:18]   loglikelihood: 32000/50996 (4.2 req/s, ETA 74.5min)
+[18:13:59]   loglikelihood: 32400/50996 (4.2 req/s, ETA 73.0min)
+[18:15:42]   loglikelihood: 32800/50996 (4.2 req/s, ETA 71.5min)
+[18:17:25]   loglikelihood: 33200/50996 (4.2 req/s, ETA 70.0min)
+[18:19:08]   loglikelihood: 33600/50996 (4.2 req/s, ETA 68.5min)
+[18:20:51]   loglikelihood: 34000/50996 (4.2 req/s, ETA 67.0min)
+[18:22:34]   loglikelihood: 34400/50996 (4.2 req/s, ETA 65.5min)
+[18:24:18]   loglikelihood: 34800/50996 (4.2 req/s, ETA 64.0min)
+[18:26:02]   loglikelihood: 35200/50996 (4.2 req/s, ETA 62.5min)
+[18:27:46]   loglikelihood: 35600/50996 (4.2 req/s, ETA 61.0min)
+[18:29:31]   loglikelihood: 36000/50996 (4.2 req/s, ETA 59.5min)
+[18:31:16]   loglikelihood: 36400/50996 (4.2 req/s, ETA 57.9min)
+[18:33:02]   loglikelihood: 36800/50996 (4.2 req/s, ETA 56.4min)
+[18:34:48]   loglikelihood: 37200/50996 (4.2 req/s, ETA 54.9min)
+[18:36:35]   loglikelihood: 37600/50996 (4.2 req/s, ETA 53.4min)
+[18:38:21]   loglikelihood: 38000/50996 (4.2 req/s, ETA 51.8min)
+[18:40:08]   loglikelihood: 38400/50996 (4.2 req/s, ETA 50.3min)
+[18:41:56]   loglikelihood: 38800/50996 (4.2 req/s, ETA 48.8min)
+[18:43:44]   loglikelihood: 39200/50996 (4.2 req/s, ETA 47.2min)
+[18:45:33]   loglikelihood: 39600/50996 (4.2 req/s, ETA 45.7min)
+[18:47:22]   loglikelihood: 40000/50996 (4.2 req/s, ETA 44.1min)
+[18:49:12]   loglikelihood: 40400/50996 (4.1 req/s, ETA 42.6min)
+[18:51:03]   loglikelihood: 40800/50996 (4.1 req/s, ETA 41.1min)
+[18:52:55]   loglikelihood: 41200/50996 (4.1 req/s, ETA 39.5min)
+[18:54:47]   loglikelihood: 41600/50996 (4.1 req/s, ETA 37.9min)
+[18:56:41]   loglikelihood: 42000/50996 (4.1 req/s, ETA 36.4min)
+[18:58:37]   loglikelihood: 42400/50996 (4.1 req/s, ETA 34.8min)
+[19:00:34]   loglikelihood: 42800/50996 (4.1 req/s, ETA 33.3min)
+[19:02:34]   loglikelihood: 43200/50996 (4.1 req/s, ETA 31.7min)
+[19:04:36]   loglikelihood: 43600/50996 (4.1 req/s, ETA 30.2min)
+[19:06:50]   loglikelihood: 44000/50996 (4.1 req/s, ETA 28.6min)
+[19:10:44]   loglikelihood: 44400/50996 (4.0 req/s, ETA 27.3min)
+[19:14:58]   loglikelihood: 44800/50996 (4.0 req/s, ETA 26.0min)
+[19:19:26]   loglikelihood: 45200/50996 (3.9 req/s, ETA 24.7min)
+[19:24:06]   loglikelihood: 45600/50996 (3.9 req/s, ETA 23.3min)
+[19:28:54]   loglikelihood: 46000/50996 (3.8 req/s, ETA 22.0min)
+[19:33:50]   loglikelihood: 46400/50996 (3.7 req/s, ETA 20.5min)
+[19:38:43]   loglikelihood: 46800/50996 (3.7 req/s, ETA 19.0min)
+[19:43:36]   loglikelihood: 47200/50996 (3.6 req/s, ETA 17.4min)
+[19:48:29]   loglikelihood: 47600/50996 (3.6 req/s, ETA 15.8min)
+[19:53:22]   loglikelihood: 48000/50996 (3.5 req/s, ETA 14.1min)
+[19:58:15]   loglikelihood: 48400/50996 (3.5 req/s, ETA 12.4min)
+[20:03:08]   loglikelihood: 48800/50996 (3.4 req/s, ETA 10.6min)
+[20:08:01]   loglikelihood: 49200/50996 (3.4 req/s, ETA 8.8min)
+[20:12:53]   loglikelihood: 49600/50996 (3.4 req/s, ETA 6.9min)
+[20:17:46]   loglikelihood: 50000/50996 (3.3 req/s, ETA 5.0min)
+[20:22:39]   loglikelihood: 50400/50996 (3.3 req/s, ETA 3.0min)
+[20:27:32]   loglikelihood: 50800/50996 (3.2 req/s, ETA 1.0min)
+[20:29:56]   loglikelihood done: 50996 in 15789s (3.2 req/s)
+[20:30:07] 
+All tasks completed in 15881s
+[20:30:08] Saved: /dev/shm/eval/results_mc/full_results.json
+[20:30:08] 
+  polish_psc_multiple_choice:
+[20:30:08]     acc,none: 0.9657
+[20:30:08]     acc_stderr,none: 0.0055
+[20:30:08]     f1,none: 0.9423
+[20:30:08]     acc_norm,none: 0.9657
+[20:30:08]     acc_norm_stderr,none: 0.0055
+[20:30:08] 
+  polish_ppc_multiple_choice:
+[20:30:08]     acc,none: 0.7790
+[20:30:08]     acc_stderr,none: 0.0131
+[20:30:08]     acc_norm,none: 0.7790
+[20:30:08]     acc_norm_stderr,none: 0.0131
+[20:30:08] 
+  polish_cbd_multiple_choice:
+[20:30:08]     acc,none: 0.7250
+[20:30:08]     acc_stderr,none: 0.0141
+[20:30:08]     f1,none: 0.2691
+[20:30:08]     acc_norm,none: 0.8140
+[20:30:08]     acc_norm_stderr,none: 0.0123
+[20:30:08] 
+  polish_8tags_multiple_choice:
+[20:30:08]     acc,none: 0.7452
+[20:30:08]     acc_stderr,none: 0.0066
+[20:30:08]     acc_norm,none: 0.3742
+[20:30:08]     acc_norm_stderr,none: 0.0073
+[20:30:08] 
+  polemo2_out_multiple_choice:
+[20:30:08]     acc,none: 0.7449
+[20:30:08]     acc_stderr,none: 0.0196
+[20:30:08]     acc_norm,none: 0.3300
+[20:30:08]     acc_norm_stderr,none: 0.0212
+[20:30:08] 
+  polemo2_in_multiple_choice:
+[20:30:08]     acc,none: 0.8518
+[20:30:08]     acc_stderr,none: 0.0132
+[20:30:08]     acc_norm,none: 0.1551
+[20:30:08]     acc_norm_stderr,none: 0.0135
+[20:30:08] 
+============================================================
+[20:30:08]   FINAL RESULTS SUMMARY
+[20:30:08] ============================================================
+Traceback (most recent call last):
+  File "/dev/shm/eval/eval_polish_quip.py", line 438, in <module>
+    main()
+  File "/dev/shm/eval/eval_polish_quip.py", line 415, in main
+    if key in metrics:
+       ^^^^^^^^^^^^^^
+TypeError: argument of type 'float' is not iterable
+=== RUN 2: Generate tasks ===
+I0221 20:30:11.794430 11364 utils.py:148] Note: detected 192 virtual cores but NumExpr set to maximum of 64, check "NUMEXPR_MAX_THREADS" environment variable.
+I0221 20:30:11.794530 11364 utils.py:151] Note: NumExpr detected 192 cores but "NUMEXPR_MAX_THREADS" not set, so enforcing safe limit of 16.
+I0221 20:30:11.794567 11364 utils.py:164] NumExpr defaulting to 16 threads.
+I0221 20:30:11.876986 11364 config.py:58] PyTorch version 2.10.0+cu126 available.
+W0221 20:30:12.004168 11364 warnings.py:112] /dev/shm/eval/quip-sharp/lib/codebook/__init__.py:6: FutureWarning: `torch.library.impl_abstract` was renamed to `torch.library.register_fake`. Please use that instead; we will remove `torch.library.impl_abstract` in a future version of PyTorch.
+  @torch.library.impl_abstract("quip_lib::decode_matvec_e8p")
+
+W0221 20:30:12.041492 11364 warnings.py:112] /dev/shm/eval/quip-sharp/lib/codebook/__init__.py:25: FutureWarning: `torch.library.impl_abstract` was renamed to `torch.library.register_fake`. Please use that instead; we will remove `torch.library.impl_abstract` in a future version of PyTorch.
+  @torch.library.impl_abstract("quip_lib::decompress_packed_e8p")
+
+W0221 20:30:12.202058 11364 warnings.py:112] /dev/shm/eval/quip-sharp/lib/utils/matmul_had.py:96: FutureWarning: `torch.library.impl_abstract` was renamed to `torch.library.register_fake`. Please use that instead; we will remove `torch.library.impl_abstract` in a future version of PyTorch.
+  @torch.library.impl_abstract("quip_lib::hadamard")
+
+W0221 20:30:34.602747 11364 warnings.py:112] /dev/shm/eval/lm-evaluation-harness/lm_eval/filters/extraction.py:98: SyntaxWarning: invalid escape sequence '\s'
+  - step 2 : We parse the choice with regex :[\s]*([A-?]), where ? varies by number of choices.
+
+W0221 20:30:34.603000 11364 warnings.py:112] /dev/shm/eval/lm-evaluation-harness/lm_eval/filters/extraction.py:168: SyntaxWarning: invalid escape sequence '\s'
+  f":[\s]*({without_paren_fallback_regex})"
+
+[20:30:34] ============================================================
+[20:30:34]   Polish LLM Leaderboard Eval
+[20:30:34]   Model: QuIP# Bielik-Q2-Sharp Variant A
+[20:30:34]   lm-eval API: new
+[20:30:34]   Tasks: ['polemo2_in', 'polemo2_out', 'polish_8tags_regex', 'polish_cbd_regex', 'polish_ppc_regex', 'polish_psc_regex']
+[20:30:34]   Few-shot: 5
+[20:30:34] ============================================================
+[20:30:34] Loading QuIP# model from /dev/shm/eval/model...
+I0221 20:30:36.434183 11364 modeling.py:987] We will use 90% of the memory on device 0 for storing the model, and 10% for the buffer to avoid OOM. You can set `max_memory` in to a higher value to use more memory (at your own risk).
+[20:30:37] Model loaded in 3.1s
+[20:30:38] Tokenizer: speakleash/Bielik-11B-v2.3-Instruct (vocab=32000)
+[20:30:38] 
+Running 6 tasks...
+I0221 20:30:38.052223 11364 evaluator.py:152] Setting random seed to 0 | Setting numpy seed to 1234 | Setting torch manual seed to 1234
+I0221 20:30:38.052285 11364 evaluator.py:203] Using pre-initialized model
+W0221 20:30:52.024997 11364 evaluator.py:251] Overwriting default num_fewshot of polish_psc_regex from None to 5
+I0221 20:30:52.025111 11364 evaluator.py:261] Setting fewshot random generator seed to 1234
+W0221 20:30:52.025141 11364 evaluator.py:251] Overwriting default num_fewshot of polish_ppc_regex from None to 5
+I0221 20:30:52.025173 11364 evaluator.py:261] Setting fewshot random generator seed to 1234
+W0221 20:30:52.025190 11364 evaluator.py:251] Overwriting default num_fewshot of polish_cbd_regex from None to 5
+I0221 20:30:52.025216 11364 evaluator.py:261] Setting fewshot random generator seed to 1234
+W0221 20:30:52.025230 11364 evaluator.py:251] Overwriting default num_fewshot of polish_8tags_regex from None to 5
+I0221 20:30:52.025256 11364 evaluator.py:261] Setting fewshot random generator seed to 1234
+W0221 20:30:52.025269 11364 evaluator.py:251] Overwriting default num_fewshot of polemo2_out from None to 5
+I0221 20:30:52.025290 11364 evaluator.py:261] Setting fewshot random generator seed to 1234
+W0221 20:30:52.025303 11364 evaluator.py:251] Overwriting default num_fewshot of polemo2_in from None to 5
+I0221 20:30:52.025323 11364 evaluator.py:261] Setting fewshot random generator seed to 1234
+I0221 20:30:52.026526 11364 task.py:410] Building contexts for polish_psc_regex on rank 0...
+
  0%|          | 0/1078 [00:00<?, ?it/s]
  1%|▏         | 16/1078 [00:00<00:06, 159.86it/s]
  3%|▎         | 33/1078 [00:00<00:06, 160.35it/s]
  5%|▍         | 50/1078 [00:00<00:06, 160.67it/s]
  6%|▌         | 67/1078 [00:00<00:06, 161.17it/s]
  8%|▊         | 84/1078 [00:00<00:06, 161.53it/s]
  9%|▉         | 101/1078 [00:00<00:06, 161.66it/s]
 11%|█         | 118/1078 [00:00<00:05, 161.58it/s]
 13%|█▎        | 135/1078 [00:00<00:05, 161.34it/s]
 14%|█▍        | 152/1078 [00:00<00:05, 161.22it/s]
 16%|█▌        | 169/1078 [00:01<00:05, 161.34it/s]
 17%|█▋        | 186/1078 [00:01<00:05, 161.34it/s]
 19%|█▉        | 203/1078 [00:01<00:05, 160.76it/s]
 20%|██        | 220/1078 [00:01<00:05, 160.64it/s]
 22%|██▏       | 237/1078 [00:01<00:05, 160.79it/s]
 24%|██▎       | 254/1078 [00:01<00:05, 160.66it/s]
 25%|██▌       | 271/1078 [00:01<00:05, 160.46it/s]
 27%|██▋       | 288/1078 [00:01<00:04, 160.49it/s]
 28%|██▊       | 305/1078 [00:01<00:04, 160.39it/s]
 30%|██▉       | 322/1078 [00:02<00:04, 160.32it/s]
 31%|███▏      | 339/1078 [00:02<00:04, 160.32it/s]
 33%|███▎      | 356/1078 [00:02<00:04, 159.71it/s]
 35%|███▍      | 373/1078 [00:02<00:04, 159.80it/s]
 36%|███▌      | 389/1078 [00:02<00:04, 159.70it/s]
 38%|███▊      | 405/1078 [00:02<00:04, 159.72it/s]
 39%|███▉      | 421/1078 [00:02<00:04, 159.58it/s]
 41%|████      | 438/1078 [00:02<00:04, 159.76it/s]
 42%|████▏     | 454/1078 [00:02<00:03, 159.47it/s]
 44%|████▎     | 470/1078 [00:02<00:03, 159.44it/s]
 45%|████▌     | 486/1078 [00:03<00:03, 159.47it/s]
 47%|████▋     | 502/1078 [00:03<00:03, 159.47it/s]
 48%|████▊     | 518/1078 [00:03<00:03, 159.31it/s]
 50%|████▉     | 534/1078 [00:03<00:03, 159.25it/s]
 51%|█████     | 550/1078 [00:03<00:03, 159.35it/s]
 53%|█████▎    | 566/1078 [00:03<00:03, 159.43it/s]
 54%|█████▍    | 582/1078 [00:03<00:03, 159.34it/s]
 55%|█████▌    | 598/1078 [00:03<00:03, 159.14it/s]
 57%|█████▋    | 614/1078 [00:03<00:02, 159.22it/s]
 58%|█████▊    | 630/1078 [00:03<00:02, 159.19it/s]
 60%|█████▉    | 646/1078 [00:04<00:02, 159.05it/s]
 61%|██████▏   | 662/1078 [00:04<00:02, 159.21it/s]
 63%|██████▎   | 678/1078 [00:04<00:02, 159.19it/s]
 64%|██████▍   | 694/1078 [00:04<00:02, 158.89it/s]
 66%|██████▌   | 710/1078 [00:04<00:02, 158.85it/s]
 67%|██████▋   | 726/1078 [00:04<00:02, 159.02it/s]
 69%|██████▉   | 742/1078 [00:04<00:02, 159.25it/s]
 70%|███████   | 758/1078 [00:04<00:02, 157.28it/s]
 72%|███████▏  | 774/1078 [00:04<00:01, 157.86it/s]
 73%|███████▎  | 790/1078 [00:04<00:01, 158.46it/s]
 75%|███████▍  | 806/1078 [00:05<00:01, 158.75it/s]
 76%|███████▋  | 822/1078 [00:05<00:01, 159.04it/s]
 78%|███████▊  | 838/1078 [00:05<00:01, 159.06it/s]
 79%|███████▉  | 854/1078 [00:05<00:01, 159.15it/s]
 81%|████████  | 870/1078 [00:05<00:01, 159.29it/s]
 82%|████████▏ | 886/1078 [00:05<00:01, 159.35it/s]
 84%|████████▎ | 902/1078 [00:05<00:01, 158.12it/s]
 85%|████████▌ | 918/1078 [00:05<00:01, 158.24it/s]
 87%|████████▋ | 934/1078 [00:05<00:00, 158.38it/s]
 88%|████████▊ | 950/1078 [00:05<00:00, 156.43it/s]
 90%|████████▉ | 966/1078 [00:06<00:00, 157.20it/s]
 91%|█████████ | 982/1078 [00:06<00:00, 157.84it/s]
 93%|█████████▎| 998/1078 [00:06<00:00, 155.60it/s]
 94%|█████████▍| 1014/1078 [00:06<00:00, 156.56it/s]
 96%|█████████▌| 1030/1078 [00:06<00:00, 154.71it/s]
 97%|█████████▋| 1046/1078 [00:06<00:00, 155.64it/s]
 99%|█████████▊| 1062/1078 [00:06<00:00, 153.76it/s]
+I0221 20:30:58.827300 11364 task.py:410] Building contexts for polish_ppc_regex on rank 0...
+
  0%|          | 0/1000 [00:00<?, ?it/s]
  2%|▏         | 17/1000 [00:00<00:06, 160.20it/s]
  3%|▎         | 34/1000 [00:00<00:05, 165.63it/s]
  5%|▌         | 51/1000 [00:00<00:06, 143.61it/s]
  7%|▋         | 66/1000 [00:00<00:06, 139.27it/s]
  8%|▊         | 81/1000 [00:00<00:06, 141.97it/s]
 10%|▉         | 97/1000 [00:00<00:06, 147.44it/s]
 11%|█▏        | 114/1000 [00:00<00:05, 154.03it/s]
 13%|█▎        | 130/1000 [00:00<00:05, 155.63it/s]
 15%|█▍        | 147/1000 [00:00<00:05, 159.40it/s]
 16%|█▋        | 164/1000 [00:01<00:05, 161.40it/s]
 18%|█▊        | 181/1000 [00:01<00:05, 162.45it/s]
 20%|█▉        | 198/1000 [00:01<00:04, 163.69it/s]
 22%|██▏       | 215/1000 [00:01<00:04, 164.83it/s]
 23%|██▎       | 232/1000 [00:01<00:04, 165.63it/s]
 25%|██▍       | 249/1000 [00:01<00:05, 138.25it/s]
 27%|██▋       | 266/1000 [00:01<00:05, 146.07it/s]
 28%|██▊       | 283/1000 [00:01<00:04, 152.15it/s]
 30%|███       | 300/1000 [00:01<00:04, 155.97it/s]
 32%|███▏      | 317/1000 [00:02<00:04, 159.29it/s]
 33%|███▎      | 334/1000 [00:02<00:04, 160.91it/s]
 35%|███▌      | 351/1000 [00:02<00:03, 162.78it/s]
 37%|███▋      | 368/1000 [00:02<00:03, 164.50it/s]
 38%|███▊      | 385/1000 [00:02<00:03, 165.56it/s]
 40%|████      | 402/1000 [00:02<00:03, 166.38it/s]
 42%|████▏     | 419/1000 [00:02<00:03, 167.03it/s]
 44%|████▎     | 436/1000 [00:02<00:03, 167.73it/s]
 45%|████▌     | 453/1000 [00:02<00:03, 167.94it/s]
 47%|████▋     | 470/1000 [00:02<00:03, 167.98it/s]
 49%|████▊     | 487/1000 [00:03<00:03, 168.28it/s]
 50%|█████     | 504/1000 [00:03<00:02, 168.77it/s]
 52%|█████▏    | 521/1000 [00:03<00:04, 116.17it/s]
 54%|█████▎    | 535/1000 [00:03<00:07, 62.33it/s] 
 55%|█████▌    | 553/1000 [00:04<00:05, 78.42it/s]
 57%|█████▋    | 570/1000 [00:04<00:05, 83.60it/s]
 58%|█████▊    | 582/1000 [00:04<00:04, 88.18it/s]
 59%|█████▉    | 594/1000 [00:04<00:04, 87.09it/s]
 61%|██████    | 612/1000 [00:04<00:03, 105.05it/s]
 63%|██████▎   | 630/1000 [00:04<00:03, 120.43it/s]
 65%|██████▍   | 648/1000 [00:04<00:02, 132.97it/s]
 66%|██████▋   | 665/1000 [00:04<00:02, 142.25it/s]
 68%|██████▊   | 682/1000 [00:04<00:02, 149.57it/s]
 70%|███████   | 700/1000 [00:05<00:01, 155.55it/s]
 72%|███████▏  | 718/1000 [00:05<00:01, 159.84it/s]
 74%|███████▎  | 736/1000 [00:05<00:01, 162.98it/s]
 75%|███████▌  | 753/1000 [00:05<00:01, 164.88it/s]
 77%|███████▋  | 770/1000 [00:05<00:01, 165.84it/s]
 79%|███████▉  | 788/1000 [00:05<00:01, 167.21it/s]
 81%|████████  | 806/1000 [00:05<00:01, 168.03it/s]
 82%|████████▏ | 823/1000 [00:05<00:01, 168.29it/s]
 84%|████████▍ | 840/1000 [00:05<00:00, 168.27it/s]
 86%|████████▌ | 858/1000 [00:06<00:00, 168.84it/s]
 88%|████████▊ | 875/1000 [00:06<00:00, 169.09it/s]
 89%|████████▉ | 893/1000 [00:06<00:00, 169.49it/s]
 91%|█████████ | 911/1000 [00:06<00:00, 169.94it/s]
 93%|█████████▎| 929/1000 [00:06<00:00, 170.03it/s]
 95%|█████████▍| 947/1000 [00:06<00:00, 170.27it/s]
 96%|█████████▋| 965/1000 [00:06<00:00, 170.50it/s]
 98%|█████████▊| 983/1000 [00:06<00:00, 170.60it/s]
+I0221 20:31:05.696070 11364 task.py:410] Building contexts for polish_cbd_regex on rank 0...
+
  0%|          | 0/1000 [00:00<?, ?it/s]
  2%|▏         | 19/1000 [00:00<00:05, 188.41it/s]
  4%|▍         | 38/1000 [00:00<00:05, 188.70it/s]
  6%|▌         | 57/1000 [00:00<00:05, 187.85it/s]
  8%|▊         | 76/1000 [00:00<00:04, 187.95it/s]
 10%|▉         | 95/1000 [00:00<00:04, 187.77it/s]
 11%|█▏        | 114/1000 [00:00<00:04, 186.91it/s]
 13%|█▎        | 133/1000 [00:00<00:04, 187.07it/s]
 15%|█▌        | 152/1000 [00:00<00:04, 187.30it/s]
 17%|█▋        | 171/1000 [00:00<00:04, 187.27it/s]
 19%|█▉        | 190/1000 [00:01<00:04, 186.96it/s]
 21%|██        | 209/1000 [00:01<00:04, 186.51it/s]
 23%|██▎       | 228/1000 [00:01<00:04, 186.92it/s]
 25%|██▍       | 247/1000 [00:01<00:04, 187.10it/s]
 27%|██▋       | 266/1000 [00:01<00:03, 186.76it/s]
 28%|██▊       | 285/1000 [00:01<00:03, 186.02it/s]
 30%|███       | 304/1000 [00:01<00:03, 186.02it/s]
 32%|███▏      | 323/1000 [00:01<00:03, 186.49it/s]
 34%|███▍      | 342/1000 [00:01<00:04, 139.84it/s]
 36%|███▌      | 358/1000 [00:02<00:06, 97.11it/s] 
 37%|███▋      | 371/1000 [00:02<00:06, 95.70it/s]
 38%|███▊      | 385/1000 [00:02<00:06, 96.13it/s]
 40%|████      | 404/1000 [00:02<00:05, 114.51it/s]
 42%|████▏     | 418/1000 [00:02<00:05, 99.40it/s] 
 43%|████▎     | 434/1000 [00:02<00:05, 111.85it/s]
 45%|████▌     | 451/1000 [00:03<00:04, 124.12it/s]
 47%|████▋     | 468/1000 [00:03<00:03, 134.85it/s]
 48%|████▊     | 485/1000 [00:03<00:03, 143.56it/s]
 50%|█████     | 502/1000 [00:03<00:03, 150.51it/s]
 52%|█████▏    | 521/1000 [00:03<00:02, 159.98it/s]
 54%|█████▍    | 540/1000 [00:03<00:02, 167.85it/s]
 56%|█████▌    | 559/1000 [00:03<00:02, 173.79it/s]
 58%|█████▊    | 578/1000 [00:03<00:02, 178.07it/s]
 60%|█████▉    | 597/1000 [00:03<00:02, 180.83it/s]
 62%|██████▏   | 616/1000 [00:03<00:02, 182.98it/s]
 64%|██████▎   | 635/1000 [00:04<00:01, 183.85it/s]
 65%|██████▌   | 654/1000 [00:04<00:01, 184.96it/s]
 67%|██████▋   | 673/1000 [00:04<00:01, 185.64it/s]
 69%|██████▉   | 692/1000 [00:04<00:01, 186.37it/s]
 71%|███████   | 711/1000 [00:04<00:01, 186.44it/s]
 73%|███████▎  | 730/1000 [00:04<00:01, 186.94it/s]
 75%|███████▍  | 749/1000 [00:04<00:01, 187.30it/s]
 77%|███████▋  | 768/1000 [00:04<00:01, 187.17it/s]
 79%|███████▊  | 787/1000 [00:04<00:01, 187.37it/s]
 81%|████████  | 806/1000 [00:04<00:01, 187.29it/s]
 82%|████████▎ | 825/1000 [00:05<00:00, 187.11it/s]
 84%|████████▍ | 844/1000 [00:05<00:00, 186.61it/s]
 86%|████████▋ | 863/1000 [00:05<00:00, 186.83it/s]
 88%|████████▊ | 882/1000 [00:05<00:00, 187.50it/s]
 90%|█████████ | 901/1000 [00:05<00:00, 179.51it/s]
 92%|█████████▏| 920/1000 [00:05<00:00, 177.24it/s]
 94%|█████████▍| 939/1000 [00:05<00:00, 179.25it/s]
 96%|█████████▌| 957/1000 [00:05<00:00, 175.66it/s]
 98%|█████████▊| 975/1000 [00:05<00:00, 173.90it/s]
 99%|█████████▉| 993/1000 [00:06<00:00, 172.82it/s]
+I0221 20:31:11.773777 11364 task.py:410] Building contexts for polish_8tags_regex on rank 0...
+
  0%|          | 0/4372 [00:00<?, ?it/s]
  0%|          | 17/4372 [00:00<00:25, 167.61it/s]
  1%|          | 35/4372 [00:00<00:25, 170.43it/s]
  1%|          | 53/4372 [00:00<00:25, 171.52it/s]
  2%|▏         | 71/4372 [00:00<00:25, 171.87it/s]
  2%|▏         | 89/4372 [00:00<00:24, 172.46it/s]
  2%|▏         | 107/4372 [00:00<00:24, 172.36it/s]
  3%|▎         | 125/4372 [00:00<00:24, 172.63it/s]
  3%|▎         | 143/4372 [00:00<00:24, 172.68it/s]
  4%|▎         | 161/4372 [00:00<00:24, 172.64it/s]
  4%|▍         | 179/4372 [00:01<00:37, 112.11it/s]
  4%|▍         | 193/4372 [00:01<00:50, 82.60it/s] 
  5%|▍         | 211/4372 [00:01<00:42, 98.80it/s]
  5%|▌         | 224/4372 [00:01<00:49, 83.86it/s]
  5%|▌         | 235/4372 [00:02<00:52, 78.49it/s]
  6%|▌         | 245/4372 [00:02<00:54, 76.28it/s]
  6%|▌         | 257/4372 [00:02<00:53, 76.80it/s]
  6%|▌         | 266/4372 [00:02<00:53, 76.49it/s]
  6%|▋         | 275/4372 [00:02<00:54, 74.64it/s]
  7%|▋         | 290/4372 [00:02<00:44, 91.77it/s]
  7%|▋         | 306/4372 [00:02<00:37, 107.86it/s]
  7%|▋         | 318/4372 [00:02<00:42, 94.79it/s] 
  8%|▊         | 329/4372 [00:03<01:00, 66.86it/s]
  8%|▊         | 338/4372 [00:03<01:07, 59.62it/s]
  8%|▊         | 356/4372 [00:03<00:49, 81.05it/s]
  9%|▊         | 373/4372 [00:03<00:40, 98.95it/s]
  9%|▉         | 386/4372 [00:03<00:39, 100.70it/s]
  9%|▉         | 398/4372 [00:03<00:43, 90.77it/s] 
 10%|▉         | 416/4372 [00:04<00:36, 109.70it/s]
 10%|▉         | 434/4372 [00:04<00:31, 125.46it/s]
 10%|█         | 452/4372 [00:04<00:28, 137.95it/s]
 11%|█         | 470/4372 [00:04<00:26, 147.11it/s]
 11%|█         | 488/4372 [00:04<00:25, 154.11it/s]
 12%|█▏        | 506/4372 [00:04<00:24, 159.15it/s]
 12%|█▏        | 524/4372 [00:04<00:23, 163.10it/s]
 12%|█▏        | 542/4372 [00:04<00:23, 165.66it/s]
 13%|█▎        | 560/4372 [00:04<00:22, 167.60it/s]
 13%|█▎        | 578/4372 [00:04<00:22, 169.25it/s]
 14%|█▎        | 596/4372 [00:05<00:22, 169.84it/s]
 14%|█▍        | 614/4372 [00:05<00:22, 170.69it/s]
 14%|█▍        | 632/4372 [00:05<00:21, 171.04it/s]
 15%|█▍        | 650/4372 [00:05<00:22, 168.21it/s]
 15%|█▌        | 667/4372 [00:05<00:22, 163.04it/s]
 16%|█▌        | 685/4372 [00:05<00:22, 165.31it/s]
 16%|█▌        | 702/4372 [00:05<00:22, 164.92it/s]
 16%|█▋        | 720/4372 [00:05<00:21, 166.88it/s]
 17%|█▋        | 738/4372 [00:05<00:21, 168.49it/s]
 17%|█▋        | 756/4372 [00:06<00:21, 169.73it/s]
 18%|█▊        | 774/4372 [00:06<00:21, 170.68it/s]
 18%|█▊        | 792/4372 [00:06<00:20, 171.19it/s]
 19%|█▊        | 810/4372 [00:06<00:20, 171.79it/s]
 19%|█▉        | 828/4372 [00:06<00:20, 172.15it/s]
 19%|█▉        | 846/4372 [00:06<00:20, 172.53it/s]
 20%|█▉        | 864/4372 [00:06<00:20, 172.32it/s]
 20%|██        | 882/4372 [00:06<00:20, 172.56it/s]
 21%|██        | 900/4372 [00:06<00:20, 172.70it/s]
 21%|██        | 918/4372 [00:06<00:20, 172.53it/s]
 21%|██▏       | 936/4372 [00:07<00:19, 172.60it/s]
 22%|██▏       | 954/4372 [00:07<00:19, 172.37it/s]
 22%|██▏       | 972/4372 [00:07<00:19, 171.35it/s]
 23%|██▎       | 990/4372 [00:07<00:20, 165.75it/s]
 23%|██▎       | 1007/4372 [00:07<00:21, 159.92it/s]
 23%|██▎       | 1024/4372 [00:07<00:21, 154.76it/s]
 24%|██▍       | 1040/4372 [00:07<00:21, 153.80it/s]
 24%|██▍       | 1056/4372 [00:07<00:21, 153.75it/s]
 25%|██▍       | 1072/4372 [00:07<00:21, 153.54it/s]
 25%|██▍       | 1088/4372 [00:08<00:21, 153.93it/s]
 25%|██▌       | 1104/4372 [00:08<00:21, 153.09it/s]
 26%|██▌       | 1120/4372 [00:08<00:21, 151.71it/s]
 26%|██▌       | 1136/4372 [00:08<00:21, 152.58it/s]
 26%|██▋       | 1152/4372 [00:08<00:21, 152.38it/s]
 27%|██▋       | 1168/4372 [00:08<00:20, 152.89it/s]
 27%|██▋       | 1184/4372 [00:08<00:20, 153.18it/s]
 27%|██▋       | 1200/4372 [00:08<00:20, 153.31it/s]
 28%|██▊       | 1216/4372 [00:08<00:20, 151.90it/s]
 28%|██▊       | 1232/4372 [00:09<00:20, 152.35it/s]
 29%|██▊       | 1248/4372 [00:09<00:20, 152.43it/s]
 29%|██▉       | 1264/4372 [00:09<00:20, 149.53it/s]
 29%|██▉       | 1280/4372 [00:09<00:21, 146.71it/s]
 30%|██▉       | 1295/4372 [00:09<00:20, 147.61it/s]
 30%|██▉       | 1311/4372 [00:09<00:20, 149.44it/s]
 30%|███       | 1327/4372 [00:09<00:20, 150.26it/s]
 31%|███       | 1343/4372 [00:09<00:20, 150.59it/s]
 31%|███       | 1359/4372 [00:09<00:19, 151.00it/s]
 31%|███▏      | 1375/4372 [00:09<00:19, 151.87it/s]
 32%|███▏      | 1391/4372 [00:10<00:19, 152.36it/s]
 32%|███▏      | 1407/4372 [00:10<00:19, 153.01it/s]
 33%|███▎      | 1423/4372 [00:10<00:19, 153.63it/s]
 33%|███▎      | 1439/4372 [00:10<00:19, 153.61it/s]
 33%|███▎      | 1455/4372 [00:10<00:18, 153.74it/s]
 34%|███▎      | 1471/4372 [00:10<00:18, 153.93it/s]
 34%|███▍      | 1487/4372 [00:10<00:18, 152.73it/s]
 34%|███▍      | 1503/4372 [00:10<00:18, 153.71it/s]
 35%|███▍      | 1519/4372 [00:10<00:18, 153.66it/s]
 35%|███▌      | 1535/4372 [00:11<00:18, 154.20it/s]
 35%|███▌      | 1551/4372 [00:11<00:18, 154.35it/s]
 36%|███▌      | 1567/4372 [00:11<00:18, 154.94it/s]
 36%|███▌      | 1583/4372 [00:11<00:17, 155.01it/s]
 37%|███▋      | 1599/4372 [00:11<00:17, 154.81it/s]
 37%|███▋      | 1615/4372 [00:11<00:17, 154.21it/s]
 37%|███▋      | 1632/4372 [00:11<00:17, 155.71it/s]
 38%|███▊      | 1649/4372 [00:11<00:17, 156.61it/s]
 38%|███▊      | 1665/4372 [00:11<00:17, 154.05it/s]
 38%|███▊      | 1682/4372 [00:12<00:20, 130.20it/s]
 39%|███▉      | 1698/4372 [00:12<00:19, 135.11it/s]
 39%|███▉      | 1714/4372 [00:12<00:19, 138.59it/s]
 40%|███▉      | 1731/4372 [00:12<00:18, 144.82it/s]
 40%|████      | 1749/4372 [00:12<00:17, 152.43it/s]
 40%|████      | 1767/4372 [00:12<00:16, 158.15it/s]
 41%|████      | 1785/4372 [00:12<00:16, 161.63it/s]
 41%|████      | 1803/4372 [00:12<00:15, 164.94it/s]
 42%|████▏     | 1821/4372 [00:12<00:15, 167.17it/s]
 42%|████▏     | 1839/4372 [00:12<00:15, 168.84it/s]
 42%|████▏     | 1856/4372 [00:13<00:15, 164.56it/s]
 43%|████▎     | 1873/4372 [00:13<00:15, 161.18it/s]
 43%|████▎     | 1890/4372 [00:13<00:15, 160.09it/s]
 44%|████▎     | 1907/4372 [00:13<00:15, 161.86it/s]
 44%|████▍     | 1924/4372 [00:13<00:14, 163.74it/s]
 44%|████▍     | 1942/4372 [00:13<00:14, 166.38it/s]
 45%|████▍     | 1959/4372 [00:13<00:24, 99.46it/s] 
 45%|████▌     | 1973/4372 [00:14<00:34, 69.25it/s]
 45%|████▌     | 1984/4372 [00:14<00:37, 63.59it/s]
 46%|████▌     | 1993/4372 [00:14<00:41, 57.60it/s]
 46%|████▌     | 2001/4372 [00:14<00:44, 53.41it/s]
 46%|████▌     | 2008/4372 [00:15<00:48, 49.25it/s]
 46%|████▌     | 2014/4372 [00:15<00:47, 49.59it/s]
 46%|████▌     | 2020/4372 [00:15<00:52, 44.82it/s]
 46%|████▋     | 2025/4372 [00:15<00:52, 44.65it/s]
 46%|████▋     | 2030/4372 [00:15<01:00, 38.91it/s]
 47%|████▋     | 2035/4372 [00:15<00:57, 40.64it/s]
 47%|████▋     | 2040/4372 [00:16<01:06, 35.20it/s]
 47%|████▋     | 2044/4372 [00:16<01:04, 36.17it/s]
 47%|████▋     | 2049/4372 [00:16<00:59, 38.82it/s]
 47%|████▋     | 2054/4372 [00:16<00:57, 40.63it/s]
 47%|████▋     | 2059/4372 [00:16<00:54, 42.39it/s]
 47%|████▋     | 2065/4372 [00:16<00:58, 39.19it/s]
 47%|████▋     | 2070/4372 [00:16<00:56, 41.02it/s]
 47%|████▋     | 2075/4372 [00:16<00:54, 42.43it/s]
 48%|████▊     | 2080/4372 [00:16<00:52, 43.67it/s]
 48%|████▊     | 2086/4372 [00:17<00:57, 39.68it/s]
 48%|████▊     | 2091/4372 [00:17<00:56, 40.13it/s]
 48%|████▊     | 2096/4372 [00:17<00:53, 42.19it/s]
 48%|████▊     | 2101/4372 [00:17<01:01, 36.88it/s]
 48%|████▊     | 2106/4372 [00:17<00:57, 39.72it/s]
 49%|████▊     | 2124/4372 [00:17<00:30, 73.98it/s]
 49%|████▉     | 2142/4372 [00:17<00:22, 100.40it/s]
 49%|████▉     | 2160/4372 [00:17<00:18, 120.35it/s]
 50%|████▉     | 2177/4372 [00:18<00:16, 133.63it/s]
 50%|█████     | 2194/4372 [00:18<00:15, 142.52it/s]
 51%|█████     | 2209/4372 [00:18<00:15, 143.58it/s]
 51%|█████     | 2225/4372 [00:18<00:14, 147.52it/s]
 51%|█████▏    | 2241/4372 [00:18<00:14, 150.33it/s]
 52%|█████▏    | 2257/4372 [00:18<00:14, 149.87it/s]
 52%|█████▏    | 2273/4372 [00:18<00:13, 151.39it/s]
 52%|█████▏    | 2289/4372 [00:18<00:13, 151.11it/s]
 53%|█████▎    | 2305/4372 [00:18<00:13, 151.16it/s]
 53%|█████▎    | 2321/4372 [00:18<00:13, 152.53it/s]
 53%|█████▎    | 2337/4372 [00:19<00:13, 152.41it/s]
 54%|█████▍    | 2353/4372 [00:19<00:13, 151.79it/s]
 54%|█████▍    | 2370/4372 [00:19<00:12, 154.53it/s]
 55%|█████▍    | 2386/4372 [00:19<00:12, 155.84it/s]
 55%|█████▍    | 2402/4372 [00:19<00:13, 147.81it/s]
 55%|█████▌    | 2417/4372 [00:19<00:13, 144.12it/s]
 56%|█████▌    | 2433/4372 [00:19<00:13, 148.07it/s]
 56%|█████▌    | 2451/4372 [00:19<00:12, 155.21it/s]
 56%|█████▋    | 2469/4372 [00:19<00:11, 160.28it/s]
 57%|█████▋    | 2487/4372 [00:20<00:11, 163.43it/s]
 57%|█████▋    | 2505/4372 [00:20<00:11, 166.10it/s]
 58%|█████▊    | 2523/4372 [00:20<00:11, 167.56it/s]
 58%|█████▊    | 2540/4372 [00:20<00:10, 167.76it/s]
 59%|█████▊    | 2558/4372 [00:20<00:10, 168.71it/s]
 59%|█████▉    | 2576/4372 [00:20<00:10, 169.56it/s]
 59%|█████▉    | 2593/4372 [00:20<00:10, 167.58it/s]
 60%|█████▉    | 2610/4372 [00:20<00:10, 162.14it/s]
 60%|██████    | 2627/4372 [00:20<00:11, 156.74it/s]
 60%|██████    | 2643/4372 [00:21<00:11, 154.60it/s]
 61%|██████    | 2659/4372 [00:21<00:11, 155.15it/s]
 61%|██████    | 2675/4372 [00:21<00:10, 154.29it/s]
 62%|██████▏   | 2691/4372 [00:21<00:10, 153.83it/s]
 62%|██████▏   | 2707/4372 [00:21<00:10, 151.67it/s]
 62%|██████▏   | 2723/4372 [00:21<00:11, 148.40it/s]
 63%|██████▎   | 2740/4372 [00:21<00:10, 151.48it/s]
 63%|██████▎   | 2756/4372 [00:21<00:10, 151.21it/s]
 63%|██████▎   | 2772/4372 [00:21<00:10, 152.43it/s]
 64%|██████▍   | 2788/4372 [00:21<00:10, 152.12it/s]
 64%|██████▍   | 2804/4372 [00:22<00:10, 152.51it/s]
 65%|██████▍   | 2820/4372 [00:22<00:10, 152.12it/s]
 65%|██████▍   | 2836/4372 [00:22<00:10, 152.25it/s]
 65%|██████▌   | 2852/4372 [00:22<00:10, 151.97it/s]
 66%|██████▌   | 2868/4372 [00:22<00:09, 152.35it/s]
 66%|██████▌   | 2884/4372 [00:22<00:09, 152.35it/s]
 66%|██████▋   | 2900/4372 [00:22<00:09, 153.81it/s]
 67%|██████▋   | 2916/4372 [00:22<00:09, 152.97it/s]
 67%|██████▋   | 2932/4372 [00:22<00:09, 153.30it/s]
 67%|██████▋   | 2948/4372 [00:23<00:09, 153.25it/s]
 68%|██████▊   | 2964/4372 [00:23<00:09, 153.57it/s]
 68%|██████▊   | 2980/4372 [00:23<00:09, 153.13it/s]
 69%|██████▊   | 2996/4372 [00:23<00:08, 153.17it/s]
 69%|██████▉   | 3012/4372 [00:23<00:09, 148.23it/s]
 69%|██████▉   | 3028/4372 [00:23<00:09, 149.17it/s]
 70%|██████▉   | 3043/4372 [00:23<00:08, 148.85it/s]
 70%|██████▉   | 3058/4372 [00:23<00:08, 148.98it/s]
 70%|███████   | 3073/4372 [00:23<00:08, 148.83it/s]
 71%|███████   | 3088/4372 [00:23<00:08, 148.75it/s]
 71%|███████   | 3104/4372 [00:24<00:08, 149.90it/s]
 71%|███████▏  | 3119/4372 [00:24<00:08, 148.49it/s]
 72%|███████▏  | 3134/4372 [00:24<00:08, 148.34it/s]
 72%|███████▏  | 3149/4372 [00:24<00:08, 147.79it/s]
 72%|███████▏  | 3164/4372 [00:24<00:08, 148.28it/s]
 73%|███████▎  | 3180/4372 [00:24<00:07, 149.25it/s]
 73%|███████▎  | 3195/4372 [00:24<00:07, 147.82it/s]
 73%|███████▎  | 3210/4372 [00:24<00:07, 148.46it/s]
 74%|███████▍  | 3226/4372 [00:24<00:07, 150.22it/s]
 74%|███████▍  | 3242/4372 [00:24<00:07, 150.98it/s]
 75%|███████▍  | 3258/4372 [00:25<00:07, 150.45it/s]
 75%|███████▍  | 3274/4372 [00:25<00:07, 150.30it/s]
 75%|███████▌  | 3290/4372 [00:25<00:07, 149.01it/s]
 76%|███████▌  | 3305/4372 [00:25<00:07, 149.17it/s]
 76%|███████▌  | 3321/4372 [00:25<00:06, 151.35it/s]
 76%|███████▋  | 3337/4372 [00:25<00:06, 151.68it/s]
 77%|███████▋  | 3353/4372 [00:25<00:06, 153.09it/s]
 77%|███████▋  | 3369/4372 [00:25<00:06, 154.78it/s]
 77%|███████▋  | 3385/4372 [00:25<00:06, 154.45it/s]
 78%|███████▊  | 3401/4372 [00:26<00:06, 152.38it/s]
 78%|███████▊  | 3417/4372 [00:26<00:06, 152.88it/s]
 79%|███████▊  | 3433/4372 [00:26<00:06, 148.80it/s]
 79%|███████▉  | 3449/4372 [00:26<00:06, 150.90it/s]
 79%|███████▉  | 3465/4372 [00:26<00:05, 151.56it/s]
 80%|███████▉  | 3481/4372 [00:26<00:05, 153.10it/s]
 80%|███████▉  | 3497/4372 [00:26<00:05, 154.35it/s]
 80%|████████  | 3513/4372 [00:26<00:05, 154.61it/s]
 81%|████████  | 3529/4372 [00:26<00:05, 155.69it/s]
 81%|████████  | 3545/4372 [00:26<00:05, 154.95it/s]
 81%|████████▏ | 3561/4372 [00:27<00:05, 153.81it/s]
 82%|████████▏ | 3577/4372 [00:27<00:05, 153.11it/s]
 82%|████████▏ | 3593/4372 [00:27<00:05, 154.18it/s]
 83%|████████▎ | 3609/4372 [00:27<00:05, 143.34it/s]
 83%|████████▎ | 3624/4372 [00:27<00:05, 133.81it/s]
 83%|████████▎ | 3642/4372 [00:27<00:05, 144.37it/s]
 84%|████████▎ | 3660/4372 [00:27<00:04, 152.37it/s]
 84%|████████▍ | 3678/4372 [00:27<00:04, 158.22it/s]
 85%|████████▍ | 3696/4372 [00:27<00:04, 162.18it/s]
 85%|████████▍ | 3714/4372 [00:28<00:03, 165.47it/s]
 85%|████████▌ | 3732/4372 [00:28<00:03, 167.76it/s]
 86%|████████▌ | 3750/4372 [00:28<00:03, 168.83it/s]
 86%|████████▌ | 3768/4372 [00:28<00:03, 169.85it/s]
 87%|████████▋ | 3786/4372 [00:28<00:03, 170.90it/s]
 87%|████████▋ | 3804/4372 [00:28<00:03, 171.44it/s]
 87%|████████▋ | 3822/4372 [00:28<00:03, 171.59it/s]
 88%|████████▊ | 3840/4372 [00:28<00:03, 172.10it/s]
 88%|████████▊ | 3858/4372 [00:28<00:02, 172.25it/s]
 89%|████████▊ | 3876/4372 [00:28<00:02, 172.61it/s]
 89%|████████▉ | 3894/4372 [00:29<00:02, 172.26it/s]
 89%|████████▉ | 3912/4372 [00:29<00:02, 172.51it/s]
 90%|████████▉ | 3930/4372 [00:29<00:02, 172.72it/s]
 90%|█████████ | 3948/4372 [00:29<00:02, 172.13it/s]
 91%|█████████ | 3966/4372 [00:29<00:02, 172.48it/s]
 91%|█████████ | 3984/4372 [00:29<00:02, 172.40it/s]
 92%|█████████▏| 4002/4372 [00:29<00:02, 172.64it/s]
 92%|█████████▏| 4020/4372 [00:29<00:02, 172.37it/s]
 92%|█████████▏| 4038/4372 [00:29<00:01, 172.64it/s]
 93%|█████████▎| 4056/4372 [00:30<00:01, 171.59it/s]
 93%|█████████▎| 4074/4372 [00:30<00:01, 171.82it/s]
 94%|█████████▎| 4092/4372 [00:30<00:01, 172.24it/s]
 94%|█████████▍| 4110/4372 [00:30<00:01, 172.41it/s]
 94%|█████████▍| 4128/4372 [00:30<00:01, 172.44it/s]
 95%|█████████▍| 4146/4372 [00:30<00:01, 172.41it/s]
 95%|█████████▌| 4164/4372 [00:30<00:01, 172.53it/s]
 96%|█████████▌| 4182/4372 [00:30<00:01, 172.57it/s]
 96%|█████████▌| 4200/4372 [00:30<00:00, 172.62it/s]
 96%|█████████▋| 4218/4372 [00:30<00:00, 172.57it/s]
 97%|█████████▋| 4236/4372 [00:31<00:00, 172.55it/s]
 97%|█████████▋| 4254/4372 [00:31<00:00, 172.68it/s]
 98%|█████████▊| 4272/4372 [00:31<00:00, 172.40it/s]
 98%|█████████▊| 4290/4372 [00:31<00:00, 172.57it/s]
 99%|█████████▊| 4308/4372 [00:31<00:00, 172.59it/s]
 99%|█████████▉| 4326/4372 [00:31<00:00, 172.58it/s]
 99%|█████████▉| 4344/4372 [00:31<00:00, 172.62it/s]
+I0221 20:31:43.701277 11364 task.py:410] Building contexts for polemo2_out on rank 0...
+
  0%|          | 0/494 [00:00<?, ?it/s]
  4%|▍         | 21/494 [00:00<00:02, 203.23it/s]
  9%|▊         | 42/494 [00:00<00:02, 201.51it/s]
 13%|█▎        | 63/494 [00:00<00:02, 202.18it/s]
 17%|█▋        | 84/494 [00:00<00:02, 202.35it/s]
 21%|██▏       | 105/494 [00:00<00:01, 202.59it/s]
 26%|██▌       | 126/494 [00:00<00:01, 202.01it/s]
 30%|██▉       | 147/494 [00:00<00:01, 202.40it/s]
 34%|███▍      | 168/494 [00:00<00:01, 202.71it/s]
 38%|███▊      | 189/494 [00:00<00:01, 202.52it/s]
 43%|████▎     | 210/494 [00:01<00:01, 202.89it/s]
 47%|████▋     | 231/494 [00:01<00:01, 203.11it/s]
 51%|█████     | 252/494 [00:01<00:01, 203.45it/s]
 55%|█████▌    | 273/494 [00:01<00:01, 203.02it/s]
 60%|█████▉    | 294/494 [00:01<00:00, 203.21it/s]
 64%|██████▍   | 315/494 [00:01<00:00, 203.60it/s]
 68%|██████▊   | 336/494 [00:01<00:00, 203.75it/s]
 72%|███████▏  | 357/494 [00:02<00:01, 115.13it/s]
 77%|███████▋  | 378/494 [00:02<00:00, 132.01it/s]
 81%|████████  | 398/494 [00:02<00:00, 146.43it/s]
 85%|████████▍ | 419/494 [00:02<00:00, 159.39it/s]
 89%|████████▉ | 439/494 [00:02<00:00, 169.40it/s]
 93%|█████████▎| 459/494 [00:02<00:00, 177.24it/s]
 97%|█████████▋| 479/494 [00:02<00:00, 183.34it/s]
+I0221 20:31:46.416867 11364 task.py:410] Building contexts for polemo2_in on rank 0...
+
  0%|          | 0/722 [00:00<?, ?it/s]
  3%|▎         | 20/722 [00:00<00:03, 198.70it/s]
  6%|▌         | 40/722 [00:00<00:03, 199.37it/s]
  8%|▊         | 60/722 [00:00<00:03, 198.94it/s]
 11%|█         | 80/722 [00:00<00:03, 199.00it/s]
 14%|█▍        | 100/722 [00:00<00:03, 199.07it/s]
 17%|█▋        | 120/722 [00:00<00:03, 199.12it/s]
 19%|█▉        | 140/722 [00:00<00:02, 199.29it/s]
 22%|██▏       | 160/722 [00:00<00:02, 199.49it/s]
 25%|██▍       | 180/722 [00:00<00:02, 198.66it/s]
 28%|██▊       | 200/722 [00:01<00:02, 198.98it/s]
 30%|███       | 220/722 [00:01<00:02, 199.21it/s]
 33%|███▎      | 241/722 [00:01<00:02, 199.71it/s]
 36%|███▋      | 262/722 [00:01<00:02, 200.35it/s]
 39%|███▉      | 283/722 [00:01<00:02, 200.40it/s]
 42%|████▏     | 304/722 [00:01<00:02, 200.46it/s]
 45%|████▌     | 325/722 [00:01<00:01, 200.59it/s]
 48%|████▊     | 346/722 [00:01<00:01, 201.30it/s]
 51%|█████     | 367/722 [00:01<00:01, 201.39it/s]
 54%|█████▎    | 388/722 [00:01<00:01, 201.63it/s]
 57%|█████▋    | 409/722 [00:02<00:01, 201.92it/s]
 60%|█████▉    | 430/722 [00:02<00:01, 201.98it/s]
 62%|██████▏   | 451/722 [00:02<00:01, 202.11it/s]
 65%|██████▌   | 472/722 [00:02<00:01, 202.03it/s]
 68%|██████▊   | 493/722 [00:02<00:01, 201.98it/s]
 71%|███████   | 514/722 [00:02<00:01, 202.02it/s]
 74%|███████▍  | 535/722 [00:02<00:00, 202.07it/s]
 77%|███████▋  | 556/722 [00:02<00:00, 202.09it/s]
 80%|███████▉  | 577/722 [00:02<00:00, 200.51it/s]
 83%|████████▎ | 598/722 [00:02<00:00, 201.13it/s]
 86%|████████▌ | 619/722 [00:03<00:00, 201.12it/s]
 89%|████████▊ | 640/722 [00:03<00:00, 201.78it/s]
 92%|█████████▏| 661/722 [00:03<00:00, 202.10it/s]
 94%|█████████▍| 682/722 [00:03<00:00, 202.53it/s]
 97%|█████████▋| 703/722 [00:03<00:00, 202.75it/s]
+I0221 20:31:50.021368 11364 evaluator.py:431] Running generate_until requests
+[20:31:50]   generate_until: 8666 requests, max_gen_toks=64
+Terminated
+=== BOTH RUNS COMPLETE ===

variant_a/logs/gen_log.txt

@@ -0,0 +1,124 @@
+I0221 21:28:04.266541 13794 utils.py:148] Note: detected 192 virtual cores but NumExpr set to maximum of 64, check "NUMEXPR_MAX_THREADS" environment variable.
+I0221 21:28:04.266615 13794 utils.py:151] Note: NumExpr detected 192 cores but "NUMEXPR_MAX_THREADS" not set, so enforcing safe limit of 16.
+I0221 21:28:04.266655 13794 utils.py:164] NumExpr defaulting to 16 threads.
+I0221 21:28:04.350871 13794 config.py:58] PyTorch version 2.10.0+cu126 available.
+W0221 21:28:04.481072 13794 warnings.py:112] /dev/shm/eval/quip-sharp/lib/codebook/__init__.py:6: FutureWarning: `torch.library.impl_abstract` was renamed to `torch.library.register_fake`. Please use that instead; we will remove `torch.library.impl_abstract` in a future version of PyTorch.
+  @torch.library.impl_abstract("quip_lib::decode_matvec_e8p")
+
+W0221 21:28:04.520201 13794 warnings.py:112] /dev/shm/eval/quip-sharp/lib/codebook/__init__.py:25: FutureWarning: `torch.library.impl_abstract` was renamed to `torch.library.register_fake`. Please use that instead; we will remove `torch.library.impl_abstract` in a future version of PyTorch.
+  @torch.library.impl_abstract("quip_lib::decompress_packed_e8p")
+
+W0221 21:28:04.529498 13794 warnings.py:112] /dev/shm/eval/quip-sharp/lib/utils/matmul_had.py:96: FutureWarning: `torch.library.impl_abstract` was renamed to `torch.library.register_fake`. Please use that instead; we will remove `torch.library.impl_abstract` in a future version of PyTorch.
+  @torch.library.impl_abstract("quip_lib::hadamard")
+
+[21:28:25] ============================================================
+[21:28:25]   Polish LLM Leaderboard Eval
+[21:28:25]   Model: QuIP# Bielik-Q2-Sharp Variant A
+[21:28:25]   lm-eval API: new
+[21:28:25]   Tasks: ['polemo2_in', 'polemo2_out', 'polish_8tags_regex', 'polish_cbd_regex', 'polish_ppc_regex', 'polish_psc_regex']
+[21:28:25]   Few-shot: 5
+[21:28:25] ============================================================
+[21:28:25] Loading QuIP# model from /dev/shm/eval/model...
+I0221 21:28:27.258036 13794 modeling.py:987] We will use 90% of the memory on device 0 for storing the model, and 10% for the buffer to avoid OOM. You can set `max_memory` in to a higher value to use more memory (at your own risk).
+[21:28:28] Model loaded in 3.5s
+[21:28:29] Tokenizer: speakleash/Bielik-11B-v2.3-Instruct (vocab=32000)
+[21:28:29] 
+Running 6 tasks...
+I0221 21:28:29.047860 13794 evaluator.py:152] Setting random seed to 0 | Setting numpy seed to 1234 | Setting torch manual seed to 1234
+I0221 21:28:29.047946 13794 evaluator.py:203] Using pre-initialized model
+W0221 21:28:44.459776 13794 evaluator.py:251] Overwriting default num_fewshot of polish_psc_regex from None to 5
+I0221 21:28:44.459892 13794 evaluator.py:261] Setting fewshot random generator seed to 1234
+W0221 21:28:44.459923 13794 evaluator.py:251] Overwriting default num_fewshot of polish_ppc_regex from None to 5
+I0221 21:28:44.459954 13794 evaluator.py:261] Setting fewshot random generator seed to 1234
+W0221 21:28:44.459970 13794 evaluator.py:251] Overwriting default num_fewshot of polish_cbd_regex from None to 5
+I0221 21:28:44.459993 13794 evaluator.py:261] Setting fewshot random generator seed to 1234
+W0221 21:28:44.460006 13794 evaluator.py:251] Overwriting default num_fewshot of polish_8tags_regex from None to 5
+I0221 21:28:44.460033 13794 evaluator.py:261] Setting fewshot random generator seed to 1234
+W0221 21:28:44.460046 13794 evaluator.py:251] Overwriting default num_fewshot of polemo2_out from None to 5
+I0221 21:28:44.460067 13794 evaluator.py:261] Setting fewshot random generator seed to 1234
+W0221 21:28:44.460079 13794 evaluator.py:251] Overwriting default num_fewshot of polemo2_in from None to 5
+I0221 21:28:44.460098 13794 evaluator.py:261] Setting fewshot random generator seed to 1234
+I0221 21:28:44.461167 13794 task.py:410] Building contexts for polish_psc_regex on rank 0...
+
  0%|          | 0/1078 [00:00<?, ?it/s]
  1%|▏         | 16/1078 [00:00<00:06, 158.81it/s]
  3%|▎         | 32/1078 [00:00<00:06, 155.11it/s]
  5%|▍         | 49/1078 [00:00<00:06, 157.50it/s]
  6%|▌         | 65/1078 [00:00<00:06, 155.98it/s]
  8%|▊         | 82/1078 [00:00<00:06, 157.67it/s]
  9%|▉         | 99/1078 [00:00<00:06, 158.86it/s]
 11%|█         | 115/1078 [00:00<00:06, 159.15it/s]
 12%|█▏        | 131/1078 [00:00<00:05, 158.29it/s]
 14%|█▎        | 147/1078 [00:00<00:05, 158.71it/s]
 15%|█▌        | 164/1078 [00:01<00:05, 159.16it/s]
 17%|█▋        | 180/1078 [00:01<00:05, 159.25it/s]
 18%|█▊        | 197/1078 [00:01<00:05, 159.50it/s]
 20%|█▉        | 213/1078 [00:01<00:05, 159.53it/s]
 21%|██        | 229/1078 [00:01<00:05, 158.88it/s]
 23%|██▎       | 245/1078 [00:01<00:05, 158.50it/s]
 24%|██▍       | 261/1078 [00:01<00:05, 158.72it/s]
 26%|██▌       | 277/1078 [00:01<00:05, 158.89it/s]
 27%|██▋       | 293/1078 [00:01<00:04, 158.74it/s]
 29%|██▊       | 309/1078 [00:01<00:04, 158.68it/s]
 30%|███       | 325/1078 [00:02<00:04, 158.98it/s]
 32%|███��      | 341/1078 [00:02<00:04, 158.67it/s]
 33%|███▎      | 357/1078 [00:02<00:04, 158.65it/s]
 35%|███▍      | 373/1078 [00:02<00:04, 158.62it/s]
 36%|███▌      | 389/1078 [00:02<00:04, 158.42it/s]
 38%|███▊      | 405/1078 [00:02<00:04, 158.60it/s]
 39%|███▉      | 421/1078 [00:02<00:04, 158.52it/s]
 41%|████      | 437/1078 [00:02<00:04, 158.47it/s]
 42%|████▏     | 453/1078 [00:02<00:03, 158.37it/s]
 44%|████▎     | 469/1078 [00:02<00:03, 158.40it/s]
 45%|████▍     | 485/1078 [00:03<00:03, 158.45it/s]
 46%|████▋     | 501/1078 [00:03<00:03, 158.41it/s]
 48%|████▊     | 517/1078 [00:03<00:03, 157.89it/s]
 49%|████▉     | 533/1078 [00:03<00:03, 157.91it/s]
 51%|█████     | 549/1078 [00:03<00:03, 157.89it/s]
 52%|█████▏    | 565/1078 [00:03<00:03, 158.17it/s]
 54%|█████▍    | 581/1078 [00:03<00:03, 157.97it/s]
 55%|█████▌    | 597/1078 [00:03<00:03, 157.99it/s]
 57%|█████▋    | 613/1078 [00:03<00:02, 158.45it/s]
 58%|█████▊    | 629/1078 [00:03<00:02, 158.69it/s]
 60%|█████▉    | 645/1078 [00:04<00:02, 158.88it/s]
 61%|██████▏   | 661/1078 [00:04<00:02, 159.02it/s]
 63%|██████▎   | 677/1078 [00:04<00:02, 159.04it/s]
 64%|██████▍   | 693/1078 [00:04<00:02, 159.10it/s]
 66%|██████▌   | 709/1078 [00:04<00:02, 159.26it/s]
 67%|██████▋   | 725/1078 [00:04<00:02, 159.39it/s]
 69%|██████▊   | 741/1078 [00:04<00:02, 159.29it/s]
 70%|███████   | 757/1078 [00:04<00:02, 158.85it/s]
 72%|███████▏  | 773/1078 [00:04<00:01, 158.93it/s]
 73%|███████▎  | 789/1078 [00:04<00:01, 159.13it/s]
 75%|███████▍  | 805/1078 [00:05<00:01, 159.36it/s]
 76%|███████▌  | 821/1078 [00:05<00:01, 158.99it/s]
 78%|███████▊  | 837/1078 [00:05<00:01, 158.99it/s]
 79%|███████▉  | 853/1078 [00:05<00:01, 159.27it/s]
 81%|████████  | 869/1078 [00:05<00:01, 159.20it/s]
 82%|████████▏ | 885/1078 [00:05<00:01, 159.12it/s]
 84%|████████▎ | 901/1078 [00:05<00:01, 159.28it/s]
 85%|████████▌ | 917/1078 [00:05<00:01, 158.94it/s]
 87%|████████▋ | 933/1078 [00:05<00:00, 159.26it/s]
 88%|████████▊ | 949/1078 [00:05<00:00, 159.17it/s]
 90%|████████▉ | 965/1078 [00:06<00:00, 159.36it/s]
 91%|█████████ | 981/1078 [00:06<00:00, 159.25it/s]
 93%|█████████▎| 998/1078 [00:06<00:00, 159.68it/s]
 94%|█████████▍| 1014/1078 [00:06<00:00, 159.46it/s]
 96%|█████████▌| 1031/1078 [00:06<00:00, 159.71it/s]
 97%|█████████▋| 1048/1078 [00:06<00:00, 159.83it/s]
 99%|█████████▊| 1064/1078 [00:06<00:00, 159.75it/s]
+I0221 21:28:51.269732 13794 task.py:410] Building contexts for polish_ppc_regex on rank 0...
+
  0%|          | 0/1000 [00:00<?, ?it/s]
  2%|▏         | 17/1000 [00:00<00:05, 167.85it/s]
  3%|▎         | 34/1000 [00:00<00:05, 168.56it/s]
  5%|▌         | 51/1000 [00:00<00:05, 168.33it/s]
  7%|▋         | 68/1000 [00:00<00:05, 168.33it/s]
  8%|▊         | 85/1000 [00:00<00:05, 168.62it/s]
 10%|█         | 102/1000 [00:00<00:05, 168.13it/s]
 12%|█▏        | 119/1000 [00:00<00:05, 168.51it/s]
 14%|█▎        | 136/1000 [00:00<00:05, 168.82it/s]
 15%|█▌        | 153/1000 [00:00<00:05, 169.08it/s]
 17%|█▋        | 170/1000 [00:01<00:04, 168.88it/s]
 19%|█▉        | 188/1000 [00:01<00:04, 169.33it/s]
 20%|██        | 205/1000 [00:01<00:04, 169.08it/s]
 22%|██▏       | 222/1000 [00:01<00:04, 169.30it/s]
 24%|██▍       | 239/1000 [00:01<00:04, 169.08it/s]
 26%|██▌       | 256/1000 [00:01<00:04, 169.34it/s]
 27%|██▋       | 273/1000 [00:01<00:04, 169.38it/s]
 29%|██▉       | 290/1000 [00:01<00:04, 169.54it/s]
 31%|███       | 307/1000 [00:01<00:04, 168.13it/s]
 32%|███▏      | 324/1000 [00:01<00:04, 167.48it/s]
 34%|███▍      | 341/1000 [00:02<00:04, 160.71it/s]
 36%|███▌      | 358/1000 [00:02<00:03, 161.61it/s]
 38%|███▊      | 375/1000 [00:02<00:03, 163.59it/s]
 39%|███▉      | 392/1000 [00:02<00:03, 165.30it/s]
 41%|████      | 409/1000 [00:02<00:03, 166.33it/s]
 43%|████▎     | 426/1000 [00:02<00:03, 166.88it/s]
 44%|████▍     | 443/1000 [00:02<00:03, 167.24it/s]
 46%|████▌     | 460/1000 [00:02<00:03, 167.75it/s]
 48%|████▊     | 477/1000 [00:02<00:03, 167.88it/s]
 49%|████▉     | 494/1000 [00:02<00:03, 168.33it/s]
 51%|█████     | 511/1000 [00:03<00:02, 167.90it/s]
 53%|█████▎    | 528/1000 [00:03<00:02, 167.95it/s]
 55%|█████▍    | 545/1000 [00:03<00:02, 168.29it/s]
 56%|█████▌    | 562/1000 [00:03<00:02, 168.74it/s]
 58%|█████▊    | 579/1000 [00:03<00:02, 168.94it/s]
 60%|█████▉    | 597/1000 [00:03<00:02, 169.32it/s]
 62%|██████▏   | 615/1000 [00:03<00:02, 169.57it/s]
 63%|██████▎   | 633/1000 [00:03<00:02, 169.80it/s]
 65%|██████▌   | 651/1000 [00:03<00:02, 169.98it/s]
 67%|██████▋   | 668/1000 [00:03<00:01, 169.91it/s]
 69%|██████▊   | 686/1000 [00:04<00:01, 170.13it/s]
 70%|███████   | 704/1000 [00:04<00:01, 170.08it/s]
 72%|███████▏  | 722/1000 [00:04<00:01, 170.21it/s]
 74%|███████▍  | 740/1000 [00:04<00:01, 169.98it/s]
 76%|███████▌  | 757/1000 [00:04<00:01, 169.94it/s]
 78%|███████▊  | 775/1000 [00:04<00:01, 170.10it/s]
 79%|███████▉  | 793/1000 [00:04<00:01, 170.04it/s]
 81%|████████  | 811/1000 [00:04<00:01, 170.08it/s]
 83%|████████▎ | 829/1000 [00:04<00:01, 169.42it/s]
 85%|████████▍ | 846/1000 [00:05<00:00, 168.92it/s]
 86%|████████▋ | 863/1000 [00:05<00:00, 163.85it/s]
 88%|████████▊ | 880/1000 [00:05<00:00, 161.14it/s]
 90%|████████▉ | 897/1000 [00:05<00:00, 159.78it/s]
 91%|█████████▏| 913/1000 [00:05<00:00, 158.02it/s]
 93%|█████████▎| 929/1000 [00:05<00:00, 156.52it/s]
 95%|█████████▍| 946/1000 [00:05<00:00, 159.59it/s]
 96%|█████████▋| 963/1000 [00:05<00:00, 162.31it/s]
 98%|█████████▊| 981/1000 [00:05<00:00, 164.77it/s]
+I0221 21:28:57.268816 13794 task.py:410] Building contexts for polish_cbd_regex on rank 0...
+
  0%|          | 0/1000 [00:00<?, ?it/s]
  2%|▏         | 19/1000 [00:00<00:05, 189.15it/s]
  4%|▍         | 38/1000 [00:00<00:05, 188.24it/s]
  6%|▌         | 57/1000 [00:00<00:05, 188.58it/s]
  8%|▊         | 76/1000 [00:00<00:04, 188.92it/s]
 10%|▉         | 95/1000 [00:00<00:04, 189.29it/s]
 11%|█▏        | 114/1000 [00:00<00:04, 188.70it/s]
 13%|█▎        | 133/1000 [00:00<00:04, 188.80it/s]
 15%|█▌        | 153/1000 [00:00<00:04, 189.37it/s]
 17%|█▋        | 172/1000 [00:00<00:04, 189.52it/s]
 19%|█▉        | 192/1000 [00:01<00:04, 189.90it/s]
 21%|██        | 212/1000 [00:01<00:04, 190.07it/s]
 23%|██▎       | 232/1000 [00:01<00:04, 190.03it/s]
 25%|██▌       | 252/1000 [00:01<00:03, 190.05it/s]
 27%|██▋       | 272/1000 [00:01<00:03, 189.98it/s]
 29%|██▉       | 292/1000 [00:01<00:03, 190.07it/s]
 31%|███       | 312/1000 [00:01<00:03, 189.98it/s]
 33%|███▎      | 331/1000 [00:01<00:03, 189.70it/s]
 35%|███▌      | 350/1000 [00:01<00:03, 189.64it/s]
 37%|███▋      | 369/1000 [00:01<00:03, 189.71it/s]
 39%|███▉      | 388/1000 [00:02<00:03, 181.30it/s]
 41%|████      | 407/1000 [00:02<00:03, 179.14it/s]
 43%|████▎     | 426/1000 [00:02<00:03, 181.42it/s]
 44%|████▍     | 445/1000 [00:02<00:03, 183.48it/s]
 46%|████▋     | 464/1000 [00:02<00:03, 178.57it/s]
 48%|████▊     | 482/1000 [00:02<00:02, 176.39it/s]
 50%|█████     | 500/1000 [00:02<00:02, 175.65it/s]
 52%|█████▏    | 518/1000 [00:02<00:02, 174.43it/s]
 54%|█████▎    | 536/1000 [00:02<00:02, 173.45it/s]
 56%|█████▌    | 556/1000 [00:03<00:02, 178.56it/s]
 57%|█████▊    | 575/1000 [00:03<00:02, 181.83it/s]
 60%|█████▉    | 595/1000 [00:03<00:02, 184.42it/s]
 62%|██████▏   | 615/1000 [00:03<00:02, 186.19it/s]
 64%|██████▎   | 635/1000 [00:03<00:01, 187.46it/s]
 66%|██████▌   | 655/1000 [00:03<00:01, 188.46it/s]
 67%|██████▋   | 674/1000 [00:03<00:01, 188.33it/s]
 69%|██████▉   | 693/1000 [00:03<00:01, 181.21it/s]
 71%|███████   | 712/1000 [00:03<00:01, 178.84it/s]
 73%|███████▎  | 730/1000 [00:03<00:01, 176.74it/s]
 75%|███████▍  | 748/1000 [00:04<00:01, 177.23it/s]
 77%|███████▋  | 767/1000 [00:04<00:01, 180.38it/s]
 79%|███████▊  | 787/1000 [00:04<00:01, 183.38it/s]
 81%|████████  | 806/1000 [00:04<00:01, 185.12it/s]
 83%|████████▎ | 826/1000 [00:04<00:00, 186.76it/s]
 84%|████████▍ | 845/1000 [00:04<00:00, 187.38it/s]
 86%|████████▋ | 864/1000 [00:04<00:00, 188.09it/s]
 88%|████████▊ | 883/1000 [00:04<00:00, 187.67it/s]
 90%|█████████ | 903/1000 [00:04<00:00, 188.36it/s]
 92%|█████████▏| 922/1000 [00:04<00:00, 188.58it/s]
 94%|█████████▍| 941/1000 [00:05<00:00, 188.76it/s]
 96%|█████████▌| 960/1000 [00:05<00:00, 181.33it/s]
 98%|█████████▊| 979/1000 [00:05<00:00, 178.92it/s]
+I0221 21:29:02.711433 13794 task.py:410] Building contexts for polish_8tags_regex on rank 0...
+
  0%|          | 0/4372 [00:00<?, ?it/s]
  0%|          | 16/4372 [00:00<00:28, 153.78it/s]
  1%|          | 32/4372 [00:00<00:27, 155.26it/s]
  1%|          | 48/4372 [00:00<00:28, 153.14it/s]
  1%|▏         | 65/4372 [00:00<00:27, 158.52it/s]
  2%|▏         | 81/4372 [00:00<00:27, 156.39it/s]
  2%|▏         | 99/4372 [00:00<00:26, 161.28it/s]
  3%|▎         | 116/4372 [00:00<00:25, 164.03it/s]
  3%|▎         | 133/4372 [00:00<00:25, 164.88it/s]
  3%|▎         | 151/4372 [00:00<00:25, 167.00it/s]
  4%|▍         | 169/4372 [00:01<00:25, 167.99it/s]
  4%|▍         | 187/4372 [00:01<00:24, 168.99it/s]
  5%|▍         | 205/4372 [00:01<00:24, 169.67it/s]
  5%|▌         | 223/4372 [00:01<00:24, 170.17it/s]
  6%|▌         | 241/4372 [00:01<00:24, 170.01it/s]
  6%|▌         | 259/4372 [00:01<00:24, 170.52it/s]
  6%|▋         | 277/4372 [00:01<00:23, 170.74it/s]
  7%|▋         | 295/4372 [00:01<00:23, 170.76it/s]
  7%|▋         | 313/4372 [00:01<00:23, 170.59it/s]
  8%|▊         | 331/4372 [00:01<00:23, 170.63it/s]
  8%|▊         | 349/4372 [00:02<00:23, 170.55it/s]
  8%|▊         | 367/4372 [00:02<00:23, 170.77it/s]
  9%|▉         | 385/4372 [00:02<00:23, 170.91it/s]
  9%|▉         | 403/4372 [00:02<00:23, 170.99it/s]
 10%|▉         | 421/4372 [00:02<00:23, 171.24it/s]
 10%|█         | 439/4372 [00:02<00:22, 171.11it/s]
 10%|█         | 457/4372 [00:02<00:22, 171.15it/s]
 11%|█         | 475/4372 [00:02<00:22, 171.27it/s]
 11%|█▏        | 493/4372 [00:02<00:23, 165.95it/s]
 12%|█▏        | 510/4372 [00:03<00:23, 165.83it/s]
 12%|█▏        | 527/4372 [00:03<00:23, 163.88it/s]
 12%|█▏        | 544/4372 [00:03<00:23, 165.43it/s]
 13%|█▎        | 561/4372 [00:03<00:23, 163.89it/s]
 13%|█▎        | 578/4372 [00:03<00:23, 159.19it/s]
 14%|█▎        | 594/4372 [00:03<00:23, 158.30it/s]
 14%|█▍        | 610/4372 [00:03<00:23, 157.11it/s]
 14%|█▍        | 626/4372 [00:03<00:24, 155.95it/s]
 15%|█▍        | 644/4372 [00:03<00:23, 160.53it/s]
 15%|█▌        | 662/4372 [00:03<00:22, 163.97it/s]
 16%|█▌        | 680/4372 [00:04<00:22, 166.37it/s]
 16%|█▌        | 698/4372 [00:04<00:21, 168.14it/s]
 16%|█▋        | 716/4372 [00:04<00:21, 169.49it/s]
 17%|█▋        | 733/4372 [00:04<00:21, 169.58it/s]
 17%|█▋        | 751/4372 [00:04<00:21, 169.80it/s]
 18%|█▊        | 769/4372 [00:04<00:21, 170.50it/s]
 18%|█▊        | 787/4372 [00:04<00:20, 170.83it/s]
 18%|█▊        | 805/4372 [00:04<00:20, 171.30it/s]
 19%|█▉        | 823/4372 [00:04<00:20, 171.35it/s]
 19%|█▉        | 841/4372 [00:05<00:20, 171.64it/s]
 20%|█▉        | 859/4372 [00:05<00:20, 172.09it/s]
 20%|██        | 877/4372 [00:05<00:20, 171.72it/s]
 20%|██        | 895/4372 [00:05<00:20, 172.12it/s]
 21%|██        | 913/4372 [00:05<00:20, 172.22it/s]
 21%|██▏       | 931/4372 [00:05<00:19, 172.24it/s]
 22%|██▏       | 949/4372 [00:05<00:19, 172.72it/s]
 22%|██▏       | 967/4372 [00:05<00:19, 172.91it/s]
 23%|██▎       | 985/4372 [00:05<00:19, 172.48it/s]
 23%|██▎       | 1003/4372 [00:05<00:19, 172.45it/s]
 23%|██▎       | 1021/4372 [00:06<00:19, 172.18it/s]
 24%|██▍       | 1039/4372 [00:06<00:20, 164.14it/s]
 24%|██▍       | 1056/4372 [00:06<00:20, 162.52it/s]
 25%|██▍       | 1074/4372 [00:06<00:19, 165.34it/s]
 25%|██▍       | 1091/4372 [00:06<00:20, 161.47it/s]
 25%|██▌       | 1108/4372 [00:06<00:20, 155.92it/s]
 26%|██▌       | 1124/4372 [00:06<00:20, 155.54it/s]
 26%|██▌       | 1142/4372 [00:06<00:20, 159.59it/s]
 26%|██▋       | 1158/4372 [00:06<00:20, 158.20it/s]
 27%|██▋       | 1176/4372 [00:07<00:19, 162.28it/s]
 27%|██▋       | 1194/4372 [00:07<00:19, 165.30it/s]
 28%|██▊       | 1212/4372 [00:07<00:18, 167.24it/s]
 28%|██▊       | 1230/4372 [00:07<00:18, 168.55it/s]
 29%|██▊       | 1248/4372 [00:07<00:18, 169.59it/s]
 29%|██▉       | 1266/4372 [00:07<00:18, 170.35it/s]
 29%|██▉       | 1284/4372 [00:07<00:18, 170.57it/s]
 30%|██▉       | 1302/4372 [00:07<00:18, 170.38it/s]
 30%|███       | 1320/4372 [00:07<00:17, 170.75it/s]
 31%|███       | 1338/4372 [00:08<00:17, 170.61it/s]
 31%|███       | 1356/4372 [00:08<00:17, 170.70it/s]
 31%|███▏      | 1374/4372 [00:08<00:17, 170.66it/s]
 32%|███▏      | 1392/4372 [00:08<00:17, 170.07it/s]
 32%|███▏      | 1410/4372 [00:08<00:17, 170.12it/s]
 33%|███▎      | 1428/4372 [00:08<00:17, 170.24it/s]
 33%|███▎      | 1446/4372 [00:08<00:17, 170.24it/s]
 33%|███▎      | 1464/4372 [00:08<00:17, 170.42it/s]
 34%|███▍      | 1482/4372 [00:08<00:16, 170.73it/s]
 34%|███▍      | 1500/4372 [00:08<00:16, 170.49it/s]
 35%|███▍      | 1518/4372 [00:09<00:16, 170.94it/s]
 35%|███▌      | 1536/4372 [00:09<00:17, 164.03it/s]
 36%|███▌      | 1553/4372 [00:09<00:17, 161.26it/s]
 36%|███▌      | 1570/4372 [00:09<00:17, 163.20it/s]
 36%|███▋      | 1588/4372 [00:09<00:16, 164.03it/s]
 37%|███▋      | 1605/4372 [00:09<00:17, 158.56it/s]
 37%|███▋      | 1621/4372 [00:09<00:17, 157.45it/s]
 37%|███▋      | 1637/4372 [00:09<00:17, 156.63it/s]
 38%|███▊      | 1653/4372 [00:09<00:17, 155.98it/s]
 38%|███▊      | 1670/4372 [00:10<00:17, 158.86it/s]
 39%|███▊      | 1688/4372 [00:10<00:16, 162.71it/s]
 39%|███▉      | 1706/4372 [00:10<00:16, 165.44it/s]
 39%|███▉      | 1724/4372 [00:10<00:15, 167.22it/s]
 40%|███▉      | 1742/4372 [00:10<00:15, 168.38it/s]
 40%|████      | 1760/4372 [00:10<00:15, 169.21it/s]
 41%|████      | 1778/4372 [00:10<00:15, 169.81it/s]
 41%|████      | 1796/4372 [00:10<00:15, 170.46it/s]
 41%|████▏     | 1814/4372 [00:10<00:14, 170.83it/s]
 42%|████▏     | 1832/4372 [00:10<00:14, 171.18it/s]
 42%|████▏     | 1850/4372 [00:11<00:14, 171.38it/s]
 43%|████▎     | 1868/4372 [00:11<00:14, 171.60it/s]
 43%|████▎     | 1886/4372 [00:11<00:14, 171.71it/s]
 44%|████▎     | 1904/4372 [00:11<00:14, 171.91it/s]
 44%|████▍     | 1922/4372 [00:11<00:14, 171.81it/s]
 44%|████▍     | 1940/4372 [00:11<00:14, 171.77it/s]
 45%|████▍     | 1958/4372 [00:11<00:14, 171.83it/s]
 45%|████▌     | 1976/4372 [00:11<00:13, 171.69it/s]
 46%|████▌     | 1994/4372 [00:11<00:13, 171.92it/s]
 46%|████▌     | 2012/4372 [00:12<00:13, 170.96it/s]
 46%|████▋     | 2030/4372 [00:12<00:13, 171.17it/s]
 47%|████▋     | 2048/4372 [00:12<00:14, 164.09it/s]
 47%|████▋     | 2065/4372 [00:12<00:14, 161.87it/s]
 48%|████▊     | 2083/4372 [00:12<00:13, 164.47it/s]
 48%|████▊     | 2100/4372 [00:12<00:13, 164.22it/s]
 48%|████▊     | 2117/4372 [00:12<00:14, 159.62it/s]
 49%|████▉     | 2133/4372 [00:12<00:14, 159.57it/s]
 49%|████▉     | 2151/4372 [00:12<00:13, 163.41it/s]
 50%|████▉     | 2168/4372 [00:12<00:13, 163.77it/s]
 50%|████▉     | 2185/4372 [00:13<00:13, 164.50it/s]
 50%|█████     | 2203/4372 [00:13<00:12, 167.03it/s]
 51%|█████     | 2221/4372 [00:13<00:12, 168.82it/s]
 51%|█████     | 2239/4372 [00:13<00:12, 170.02it/s]
 52%|█████▏    | 2257/4372 [00:13<00:12, 170.67it/s]
 52%|█████▏    | 2275/4372 [00:13<00:12, 171.38it/s]
 52%|█████▏    | 2293/4372 [00:13<00:12, 171.65it/s]
 53%|█████▎    | 2311/4372 [00:13<00:12, 171.75it/s]
 53%|█████▎    | 2329/4372 [00:13<00:11, 172.13it/s]
 54%|█████▎    | 2347/4372 [00:14<00:11, 172.15it/s]
 54%|█████▍    | 2365/4372 [00:14<00:11, 172.00it/s]
 55%|█████▍    | 2383/4372 [00:14<00:11, 172.11it/s]
 55%|█████▍    | 2401/4372 [00:14<00:11, 172.34it/s]
 55%|█████▌    | 2419/4372 [00:14<00:11, 172.61it/s]
 56%|█████▌    | 2437/4372 [00:14<00:11, 172.38it/s]
 56%|█████▌    | 2455/4372 [00:14<00:11, 172.59it/s]
 57%|█████▋    | 2473/4372 [00:14<00:11, 172.35it/s]
 57%|█████▋    | 2491/4372 [00:14<00:10, 172.58it/s]
 57%|█████▋    | 2509/4372 [00:14<00:10, 172.65it/s]
 58%|█████▊    | 2527/4372 [00:15<00:10, 169.11it/s]
 58%|█████▊    | 2545/4372 [00:15<00:10, 169.57it/s]
 59%|█████▊    | 2563/4372 [00:15<00:10, 170.31it/s]
 59%|█████▉    | 2581/4372 [00:15<00:10, 171.06it/s]
 59%|█████▉    | 2599/4372 [00:15<00:10, 171.19it/s]
 60%|█████▉    | 2617/4372 [00:15<00:10, 171.40it/s]
 60%|██████    | 2635/4372 [00:15<00:10, 171.26it/s]
 61%|██████    | 2653/4372 [00:15<00:10, 171.24it/s]
 61%|██████    | 2671/4372 [00:15<00:09, 171.25it/s]
 62%|██████▏   | 2689/4372 [00:16<00:09, 171.21it/s]
 62%|██████▏   | 2707/4372 [00:16<00:09, 171.35it/s]
 62%|██████▏   | 2725/4372 [00:16<00:09, 171.17it/s]
 63%|██████▎   | 2743/4372 [00:16<00:09, 171.12it/s]
 63%|██████▎   | 2761/4372 [00:16<00:09, 171.18it/s]
 64%|██████▎   | 2779/4372 [00:16<00:09, 171.04it/s]
 64%|██████▍   | 2797/4372 [00:16<00:09, 171.09it/s]
 64%|██████▍   | 2815/4372 [00:16<00:09, 170.34it/s]
 65%|██████▍   | 2833/4372 [00:16<00:09, 170.51it/s]
 65%|██████▌   | 2851/4372 [00:16<00:08, 170.83it/s]
 66%|██████▌   | 2869/4372 [00:17<00:08, 171.02it/s]
 66%|██████▌   | 2887/4372 [00:17<00:08, 171.12it/s]
 66%|██████▋   | 2905/4372 [00:17<00:08, 171.05it/s]
 67%|██████▋   | 2923/4372 [00:17<00:08, 171.40it/s]
 67%|██████▋   | 2941/4372 [00:17<00:08, 171.49it/s]
 68%|██████▊   | 2959/4372 [00:17<00:08, 171.54it/s]
 68%|██████▊   | 2977/4372 [00:17<00:08, 171.32it/s]
 69%|██████▊   | 2995/4372 [00:17<00:08, 171.45it/s]
 69%|██████▉   | 3013/4372 [00:17<00:07, 171.41it/s]
 69%|██████▉   | 3031/4372 [00:18<00:07, 171.38it/s]
 70%|██████▉   | 3049/4372 [00:18<00:07, 171.39it/s]
 70%|███████   | 3067/4372 [00:18<00:07, 171.25it/s]
 71%|███████   | 3085/4372 [00:18<00:07, 171.49it/s]
 71%|███████   | 3103/4372 [00:18<00:07, 171.30it/s]
 71%|███████▏  | 3121/4372 [00:18<00:07, 171.41it/s]
 72%|███████▏  | 3139/4372 [00:18<00:07, 170.52it/s]
 72%|███████▏  | 3157/4372 [00:18<00:07, 170.41it/s]
 73%|███████▎  | 3175/4372 [00:18<00:07, 170.62it/s]
 73%|███████▎  | 3193/4372 [00:18<00:06, 170.51it/s]
 73%|███████▎  | 3211/4372 [00:19<00:06, 170.74it/s]
 74%|███████▍  | 3229/4372 [00:19<00:06, 170.85it/s]
 74%|███████▍  | 3247/4372 [00:19<00:06, 171.23it/s]
 75%|███████▍  | 3265/4372 [00:19<00:06, 171.43it/s]
 75%|███████▌  | 3283/4372 [00:19<00:06, 171.19it/s]
 76%|███████▌  | 3301/4372 [00:19<00:06, 171.41it/s]
 76%|███████▌  | 3319/4372 [00:19<00:06, 171.67it/s]
 76%|███████▋  | 3337/4372 [00:19<00:06, 171.90it/s]
 77%|███████▋  | 3355/4372 [00:19<00:05, 170.78it/s]
 77%|███████▋  | 3373/4372 [00:20<00:05, 171.28it/s]
 78%|███████▊  | 3391/4372 [00:20<00:05, 171.54it/s]
 78%|███████▊  | 3409/4372 [00:20<00:05, 171.62it/s]
 78%|███████▊  | 3427/4372 [00:20<00:05, 171.91it/s]
 79%|███████▉  | 3445/4372 [00:20<00:05, 171.95it/s]
 79%|███████▉  | 3463/4372 [00:20<00:05, 171.85it/s]
 80%|███████▉  | 3481/4372 [00:20<00:05, 171.91it/s]
 80%|████████  | 3499/4372 [00:20<00:05, 172.28it/s]
 80%|████████  | 3517/4372 [00:20<00:04, 172.54it/s]
 81%|████████  | 3535/4372 [00:20<00:04, 172.51it/s]
 81%|████████▏ | 3553/4372 [00:21<00:04, 172.67it/s]
 82%|████████▏ | 3571/4372 [00:21<00:04, 172.87it/s]
 82%|████████▏ | 3589/4372 [00:21<00:04, 172.91it/s]
 83%|████████▎ | 3607/4372 [00:21<00:04, 172.68it/s]
 83%|████████▎ | 3625/4372 [00:21<00:04, 172.70it/s]
 83%|████████▎ | 3643/4372 [00:21<00:04, 172.45it/s]
 84%|████████▎ | 3661/4372 [00:21<00:04, 172.58it/s]
 84%|████████▍ | 3679/4372 [00:21<00:04, 172.76it/s]
 85%|████████▍ | 3697/4372 [00:21<00:03, 172.05it/s]
 85%|████████▍ | 3715/4372 [00:22<00:03, 171.46it/s]
 85%|████████▌ | 3733/4372 [00:22<00:03, 171.52it/s]
 86%|████████▌ | 3751/4372 [00:22<00:03, 172.04it/s]
 86%|████████▌ | 3769/4372 [00:22<00:03, 172.25it/s]
 87%|████████▋ | 3787/4372 [00:22<00:03, 172.40it/s]
 87%|████████▋ | 3805/4372 [00:22<00:03, 172.60it/s]
 87%|████████▋ | 3823/4372 [00:22<00:03, 172.36it/s]
 88%|████████▊ | 3841/4372 [00:22<00:03, 172.41it/s]
 88%|████████▊ | 3859/4372 [00:22<00:02, 172.17it/s]
 89%|████████▊ | 3877/4372 [00:22<00:02, 172.15it/s]
 89%|████████▉ | 3895/4372 [00:23<00:02, 172.20it/s]
 90%|████████▉ | 3913/4372 [00:23<00:02, 171.99it/s]
 90%|████████▉ | 3931/4372 [00:23<00:02, 172.12it/s]
 90%|█████████ | 3949/4372 [00:23<00:02, 172.16it/s]
 91%|█████████ | 3967/4372 [00:23<00:02, 172.13it/s]
 91%|█████████ | 3985/4372 [00:23<00:02, 172.10it/s]
 92%|█████████▏| 4003/4372 [00:23<00:02, 171.92it/s]
 92%|█████████▏| 4021/4372 [00:23<00:02, 171.68it/s]
 92%|█████████▏| 4039/4372 [00:23<00:01, 171.53it/s]
 93%|█████████▎| 4057/4372 [00:23<00:01, 171.58it/s]
 93%|█████████▎| 4075/4372 [00:24<00:01, 171.36it/s]
 94%|█████████▎| 4093/4372 [00:24<00:01, 171.10it/s]
 94%|█████████▍| 4111/4372 [00:24<00:01, 171.00it/s]
 94%|█████████▍| 4129/4372 [00:24<00:01, 170.85it/s]
 95%|█████████▍| 4147/4372 [00:24<00:01, 170.80it/s]
 95%|█████████▌| 4165/4372 [00:24<00:01, 170.73it/s]
 96%|█████████▌| 4183/4372 [00:24<00:01, 170.86it/s]
 96%|█████████▌| 4201/4372 [00:24<00:01, 170.66it/s]
 97%|█████████▋| 4219/4372 [00:24<00:00, 170.72it/s]
 97%|█████████▋| 4237/4372 [00:25<00:00, 170.62it/s]
 97%|█████████▋| 4255/4372 [00:25<00:00, 170.83it/s]
 98%|█████████▊| 4273/4372 [00:25<00:00, 170.97it/s]
 98%|█████████▊| 4291/4372 [00:25<00:00, 170.21it/s]
 99%|█████████▊| 4309/4372 [00:25<00:00, 170.55it/s]
 99%|█████████▉| 4327/4372 [00:25<00:00, 170.77it/s]
 99%|█████████▉| 4345/4372 [00:25<00:00, 170.96it/s]
+I0221 21:29:28.599591 13794 task.py:410] Building contexts for polemo2_out on rank 0...
+
  0%|          | 0/494 [00:00<?, ?it/s]
  4%|▍         | 21/494 [00:00<00:02, 202.84it/s]
  9%|▊         | 42/494 [00:00<00:02, 203.44it/s]
 13%|█▎        | 63/494 [00:00<00:02, 193.13it/s]
 17%|█▋        | 83/494 [00:00<00:02, 187.79it/s]
 21%|██        | 102/494 [00:00<00:02, 186.31it/s]
 25%|██▍       | 123/494 [00:00<00:01, 191.40it/s]
 29%|██▉       | 144/494 [00:00<00:01, 195.29it/s]
 33%|███▎      | 165/494 [00:00<00:01, 197.79it/s]
 38%|███▊      | 186/494 [00:00<00:01, 199.80it/s]
 42%|████▏     | 206/494 [00:01<00:01, 199.79it/s]
 46%|████▌     | 227/494 [00:01<00:01, 200.91it/s]
 50%|█████     | 248/494 [00:01<00:01, 201.63it/s]
 54%|█████▍    | 269/494 [00:01<00:01, 202.22it/s]
 59%|█████▊    | 290/494 [00:01<00:01, 202.67it/s]
 63%|██████▎   | 311/494 [00:01<00:00, 203.14it/s]
 67%|██████▋   | 332/494 [00:01<00:00, 198.07it/s]
 71%|███████▏  | 352/494 [00:02<00:01, 114.48it/s]
 76%|███████▌  | 373/494 [00:02<00:00, 131.87it/s]
 80%|███████▉  | 394/494 [00:02<00:00, 147.29it/s]
 84%|████████▍ | 415/494 [00:02<00:00, 160.29it/s]
 88%|████████▊ | 436/494 [00:02<00:00, 171.03it/s]
 93%|█████████▎| 457/494 [00:02<00:00, 179.36it/s]
 97%|█████████▋| 478/494 [00:02<00:00, 185.56it/s]
+I0221 21:29:31.342681 13794 task.py:410] Building contexts for polemo2_in on rank 0...
+
  0%|          | 0/722 [00:00<?, ?it/s]
  3%|▎         | 21/722 [00:00<00:03, 200.52it/s]
  6%|▌         | 42/722 [00:00<00:03, 201.21it/s]
  9%|▊         | 63/722 [00:00<00:03, 201.15it/s]
 12%|█▏        | 84/722 [00:00<00:03, 201.57it/s]
 15%|█▍        | 105/722 [00:00<00:03, 200.23it/s]
 17%|█▋        | 126/722 [00:00<00:02, 201.09it/s]
 20%|██        | 147/722 [00:00<00:02, 201.34it/s]
 23%|██▎       | 168/722 [00:00<00:02, 201.57it/s]
 26%|██▌       | 189/722 [00:00<00:02, 201.72it/s]
 29%|██▉       | 210/722 [00:01<00:02, 202.09it/s]
 32%|███▏      | 231/722 [00:01<00:02, 202.36it/s]
 35%|███▍      | 252/722 [00:01<00:02, 202.48it/s]
 38%|███▊      | 273/722 [00:01<00:02, 202.68it/s]
 41%|████      | 294/722 [00:01<00:02, 203.01it/s]
 44%|████▎     | 315/722 [00:01<00:02, 203.00it/s]
 47%|████▋     | 336/722 [00:01<00:01, 203.43it/s]
 49%|████▉     | 357/722 [00:01<00:01, 203.69it/s]
 52%|█████▏    | 378/722 [00:01<00:01, 203.54it/s]
 55%|█████▌    | 399/722 [00:01<00:01, 203.70it/s]
 58%|█████▊    | 420/722 [00:02<00:01, 204.11it/s]
 61%|██████    | 441/722 [00:02<00:01, 203.85it/s]
 64%|██████▍   | 462/722 [00:02<00:01, 204.05it/s]
 67%|██████▋   | 483/722 [00:02<00:01, 204.28it/s]
 70%|██████▉   | 504/722 [00:02<00:01, 204.46it/s]
 73%|███████▎  | 525/722 [00:02<00:00, 204.96it/s]
 76%|███████▌  | 546/722 [00:02<00:00, 205.16it/s]
 79%|███████▊  | 567/722 [00:02<00:00, 204.96it/s]
 81%|████████▏ | 588/722 [00:02<00:00, 205.08it/s]
 84%|████████▍ | 609/722 [00:02<00:00, 205.21it/s]
 87%|████████▋ | 630/722 [00:03<00:00, 205.67it/s]
 90%|█████████ | 651/722 [00:03<00:00, 205.68it/s]
 93%|█████████▎| 672/722 [00:03<00:00, 205.68it/s]
 96%|█████████▌| 693/722 [00:03<00:00, 205.91it/s]
 99%|█████████▉| 714/722 [00:03<00:00, 205.58it/s]
+I0221 21:29:34.897518 13794 evaluator.py:431] Running generate_until requests
+[21:29:46]   generate_until: 8666 requests, batch_size=32 (length-sorted)
+[21:29:46]   context lengths: min=668, max=1998, median=798, max_gen_toks=64
+[21:35:30]   generate_until: 320/8666 (55.8 req/min, ETA 149.5min)
+[21:41:15]   generate_until: 640/8666 (55.8 req/min, ETA 143.9min)
+[21:47:00]   generate_until: 960/8666 (55.7 req/min, ETA 138.3min)
+[21:52:45]   generate_until: 1280/8666 (55.7 req/min, ETA 132.6min)
+[21:58:34]   generate_until: 1600/8666 (55.6 req/min, ETA 127.2min)
+[22:04:22]   generate_until: 1920/8666 (55.5 req/min, ETA 121.6min)
+[22:10:11]   generate_until: 2240/8666 (55.4 req/min, ETA 116.0min)
+[22:16:00]   generate_until: 2560/8666 (55.4 req/min, ETA 110.3min)
+[22:21:49]   generate_until: 2880/8666 (55.3 req/min, ETA 104.6min)
+[22:27:42]   generate_until: 3200/8666 (55.2 req/min, ETA 98.9min)
+[22:33:33]   generate_until: 3520/8666 (55.2 req/min, ETA 93.2min)
+[22:39:24]   generate_until: 3840/8666 (55.2 req/min, ETA 87.5min)
+[22:45:18]   generate_until: 4160/8666 (55.1 req/min, ETA 81.8min)
+[22:51:12]   generate_until: 4480/8666 (55.0 req/min, ETA 76.1min)
+[22:57:11]   generate_until: 4800/8666 (54.9 req/min, ETA 70.4min)
+[23:03:12]   generate_until: 5120/8666 (54.8 req/min, ETA 64.7min)
+[23:09:13]   generate_until: 5440/8666 (54.7 req/min, ETA 59.0min)
+[23:15:17]   generate_until: 5760/8666 (54.6 req/min, ETA 53.2min)
+[23:21:26]   generate_until: 6080/8666 (54.4 req/min, ETA 47.5min)
+[23:27:51]   generate_until: 6400/8666 (54.2 req/min, ETA 41.8min)
+[23:36:06]   generate_until: 6720/8666 (53.2 req/min, ETA 36.6min)
+[23:44:22]   generate_until: 7040/8666 (52.3 req/min, ETA 31.1min)
+[23:52:44]   generate_until: 7360/8666 (51.5 req/min, ETA 25.4min)
+[00:01:05]   generate_until: 7680/8666 (50.8 req/min, ETA 19.4min)
+[00:09:28]   generate_until: 8000/8666 (50.1 req/min, ETA 13.3min)
+[00:17:48]   generate_until: 8320/8666 (49.5 req/min, ETA 7.0min)
+[00:26:13]   generate_until: 8640/8666 (49.0 req/min, ETA 0.5min)
+[00:26:59]   generate_until: 8666/8666 (48.9 req/min, ETA 0.0min)
+[00:26:59]   generate_until done: 8666 in 10633s (48.9 req/min)
+[00:27:35] 
+All tasks completed in 10747s
+[00:27:36] Saved: /dev/shm/eval/results_gen/full_results.json
+[00:27:36] 
+  polish_psc_regex:
+[00:27:36]     exact_match,score-first: 0.9536
+[00:27:36]     exact_match_stderr,score-first: 0.0064
+[00:27:36]     f1,score-first: 0.9662
+[00:27:36] 
+  polish_ppc_regex:
+[00:27:36]     exact_match,score-first: 0.7890
+[00:27:36]     exact_match_stderr,score-first: 0.0129
+[00:27:36] 
+  polish_cbd_regex:
+[00:27:36]     exact_match,score-first: 0.7570
+[00:27:36]     exact_match_stderr,score-first: 0.0136
+[00:27:36]     f1,score-first: 0.3009
+[00:27:36] 
+  polish_8tags_regex:
+[00:27:36]     exact_match,score-first: 0.7656
+[00:27:36]     exact_match_stderr,score-first: 0.0064
+[00:27:36] 
+  polemo2_out:
+[00:27:36]     exact_match,score-first: 0.7186
+[00:27:36]     exact_match_stderr,score-first: 0.0203
+[00:27:36] 
+  polemo2_in:
+[00:27:36]     exact_match,score-first: 0.8310
+[00:27:36]     exact_match_stderr,score-first: 0.0140
+[00:27:36] 
+============================================================
+[00:27:36]   FINAL RESULTS SUMMARY
+[00:27:36] ============================================================
+Traceback (most recent call last):
+  File "/dev/shm/eval/eval_polish_quip.py", line 481, in <module>
+    main()
+  File "/dev/shm/eval/eval_polish_quip.py", line 458, in main
+    if key in metrics:
+       ^^^^^^^^^^^^^^
+TypeError: argument of type 'numpy.float64' is not iterable

variant_a/report/variant_a_report.md

@@ -0,0 +1,127 @@
+# Wariant A — QuIP# E8P12 Bielik-11B-v2.3-Instruct
+
+## 1. Parametry kwantyzacji
+
+| Parametr | Wartość |
+|---|---|
+| **Base model** | speakleash/Bielik-11B-v2.3-Instruct |
+| **Codebook** | E8P12 (2-bit, lattice E8) |
+| **Scale** | default (per-channel) |
+| **Calibration** | RedPajama-1T (custom subset) |
+| **Hessian samples** | 200 layers × ~128 samples each |
+| **Epochs** | 1 (single-pass Hessian + quantization) |
+| **Batch size (Hessian)** | 1 (sequential layer processing) |
+| **Seed** | 0 |
+| **Precision** | FP16 base → 2-bit QuIP# |
+
+## 2. Rozmiar modelu
+
+| Model | Rozmiar |
+|---|---|
+| FP16 Instruct (base) | ~22 GB |
+| **QuIP# E8P12 (2-bit)** | **3.26 GB** |
+| Kompresja | **~6.7x** |
+
+## 3. Czas i compute
+
+### Hessian generation
+- **Maszyna**: NVIDIA H200 (141 GB VRAM), vast.ai
+- **Czas**: ~2-3h na 200 hessianów
+- **GPU usage**: ~95% GPU, ~10 GB VRAM
+- **Hessians**: 200 plików .pt uploaded na Jakubrd4/bielik-quip-e8p12
+
+### Kwantyzacja
+- **Maszyna**: ta sama H200
+- **Czas**: ~1-2h
+- **GPU usage**: ~100% GPU, ~25 GB VRAM peak
+- **Output**: 3.26 GB model uploaded na Jakubrd4/bielik-q2-variant-a
+
+### Ewaluacja (Polish LLM Leaderboard)
+- **Maszyna**: NVIDIA H200, vast.ai (ssh -p 12414 root@154.57.34.75)
+- **MC loglikelihood (Run 1)**: 50,996 requests, 15,789s (~4h23m), 3.2-4.6 req/s
+- **Generative regex (Run 2)**: 8,666 requests, 10,633s (~2h57m), 48.9-55.8 req/min (batch=32)
+- **Remaining 13 tasks (Run 3)**: w toku
+- **Total eval GPU cost**: ~$3-5 (H200 @ ~$2/h × ~7h)
+
+## 4. Bugi i rozwiązania
+
+### Bug 1: MistralConfig vs LlamaConfig
+- **Problem**: QuIP# zakładał LlamaConfig, ale Bielik-11B używa MistralConfig (Mistral architecture)
+- **Rozwiązanie**: Patch w `model_from_hf_path()` — konwersja MistralConfig → LlamaConfig z mapowaniem atrybutów (sliding_window → None, attention_dropout → 0)
+- **Impact**: Blokujący — bez patcha model się nie ładował
+
+### Bug 2: fast_hadamard_transform brak CUDA kernel
+- **Problem**: `fast_hadamard_transform` pakiet nie miał skompilowanego CUDA kernel na H200 (sm_90)
+- **Rozwiązanie**: Fallback na PyTorch-native Hadamard transform: `H @ x` z rekursywną konstrukcją macierzy Hadamarda
+- **Impact**: ~10-20% wolniejsze niż native CUDA, ale działa
+
+### Bug 3: Buforowany stdout w nohup
+- **Problem**: Python buforuje stdout przy nohup redirect, logi nie rosły mimo działającego procesu
+- **Rozwiązanie**: `python -u` (PYTHONUNBUFFERED=1)
+- **Impact**: Kosmetyczny, ale powodował fałszywe alarmy
+
+### Bug 4: acc_norm vs acc metryka
+- **Problem**: auto_chain.sh raportował acc_norm (56.97%) zamiast acc (80.19%), dając mylne "poniżej baseline"
+- **Rozwiązanie**: Leaderboard używa acc dla polemo2/8tags/ppc i f1 dla cbd/psc
+- **Impact**: Fałszywy negatyw — faktyczne wyniki znacznie lepsze niż raportowane
+
+### Bug 5: Sekwencyjne generate_until (~62h ETA)
+- **Problem**: generate_until przetwarzał requesty pojedynczo (2.3 req/min)
+- **Rozwiązanie**: Batched generation z left-padding i length-sorting (batch=32)
+- **Impact**: 24x przyspieszenie (2.3 → 55.8 req/min), ETA z 62h na 2.5h
+
+## 5. Wyniki eval — MC loglikelihood (Run 1)
+
+| Task | Leaderboard metric | Score |
+|---|---|---|
+| polemo2_in | acc | **0.8518** |
+| polemo2_out | acc | **0.7449** |
+| polish_8tags | acc | **0.7452** |
+| polish_cbd | f1 | **0.2691** |
+| polish_ppc | acc | **0.7790** |
+| polish_psc | f1 | **0.9423** |
+| **Średnia (leaderboard)** | | **0.7220 (72.2%)** |
+
+## 6. Wyniki eval — Generative regex (Run 2)
+
+| Task | exact_match |
+|---|---|
+| polemo2_in | **0.8310** |
+| polemo2_out | **0.7186** |
+| polish_8tags_regex | **0.7656** |
+| polish_cbd_regex | **0.7570** |
+| polish_ppc_regex | **0.7890** |
+| polish_psc_regex | **0.9536** |
+| **Średnia** | **0.8025** |
+
+## 7. Wyniki eval — Remaining 13 tasks (Run 3)
+
+*W toku — wyniki zostaną dodane po zakończeniu*
+
+## 8. Porównanie z baseline
+
+| Model | MC avg (leaderboard) | Gen avg (exact_match) |
+|---|---|---|
+| FP16 Instruct | 65.71% | — |
+| IQ2_XXS (baseline) | 61.34% | — |
+| **QuIP# E8P12** | **72.20%** | **80.25%** |
+| Delta vs IQ2_XXS | **+10.9pp** | — |
+| Delta vs FP16 | **+6.5pp** | — |
+
+## 9. Pliki i artefakty
+
+- **Model**: https://huggingface.co/Jakubrd4/bielik-q2-variant-a
+- **Hessiany**: https://huggingface.co/Jakubrd4/bielik-quip-e8p12
+- **Dokumentacja**: https://huggingface.co/Jakubrd4/bielik-q2-sharp-docs
+- **Eval skrypt**: eval_polish_quip.py (w docs repo)
+- **Wyniki MC**: results_mc/full_results.json
+- **Wyniki Gen**: results_gen/full_results.json
+- **Wyniki Remaining**: results_remaining/full_results.json (pending)
+
+## 10. Wnioski
+
+1. **QuIP# E8P12 (2-bit) znacząco bije baseline IQ2_XXS** (+10.9pp na MC leaderboard metrics)
+2. **Nawet bije FP16** na MC leaderboard (+6.5pp) — prawdopodobnie dzięki 5-shot prompting i lepszemu formatowi
+3. **CBD f1 jest słaby** (0.2691) — model generuje poprawne etykiety (acc=0.725) ale f1 jest niska (class imbalance?)
+4. **Generative exact_match** (80.25%) potwierdza silne wyniki MC
+5. **Rozmiar 3.26 GB** (6.7x kompresja) przy zachowaniu ~80% accuracy to świetny tradeoff

variant_b/config/config.json

@@ -0,0 +1,55 @@
+{
+  "_attn_implementation_autoset": true,
+  "_name_or_path": "/workspace/rotated_model",
+  "architectures": [
+    "LlamaForCausalLM"
+  ],
+  "attention_bias": false,
+  "attention_dropout": 0.0,
+  "bos_token_id": 1,
+  "eos_token_id": [
+    32001,
+    2
+  ],
+  "head_dim": 128,
+  "hidden_act": "silu",
+  "hidden_size": 4096,
+  "initializer_range": 0.02,
+  "intermediate_size": 14336,
+  "max_position_embeddings": 32768,
+  "mlp_bias": false,
+  "model_type": "llama",
+  "num_attention_heads": 32,
+  "num_hidden_layers": 50,
+  "num_key_value_heads": 8,
+  "pretraining_tp": 1,
+  "quantization_config": {
+    "bits": 2,
+    "checkpoint_format": "gptq",
+    "desc_act": false,
+    "group_size": 128,
+    "lm_head": false,
+    "meta": {
+      "damp_auto_increment": 0.0025,
+      "damp_percent": 0.01,
+      "mse": 0.0,
+      "quantizer": [
+        "gptqmodel:1.9.0"
+      ],
+      "static_groups": false,
+      "true_sequential": true,
+      "uri": "https://github.com/modelcloud/gptqmodel"
+    },
+    "pack_dtype": "int32",
+    "quant_method": "gptq",
+    "sym": true
+  },
+  "rms_norm_eps": 1e-05,
+  "rope_scaling": null,
+  "rope_theta": 1000000,
+  "tie_word_embeddings": false,
+  "torch_dtype": "bfloat16",
+  "transformers_version": "4.48.3",
+  "use_cache": true,
+  "vocab_size": 32128
+}

variant_b/config/quantization_meta.json

@@ -0,0 +1,10 @@
+{
+  "method": "SpinQuant + GPTQModel",
+  "w_bits": 2,
+  "a_bits": 16,
+  "w_groupsize": 128,
+  "desc_act": false,
+  "rotation_steps": 100,
+  "calibration": "Polish Wikipedia (128 samples)",
+  "source_model": "speakleash/Bielik-11B-v2.3-Instruct"
+}

variant_b/config/quantize_config.json

@@ -0,0 +1,21 @@
+{
+  "bits": 2,
+  "group_size": 128,
+  "desc_act": false,
+  "sym": true,
+  "lm_head": false,
+  "quant_method": "gptq",
+  "checkpoint_format": "gptq",
+  "pack_dtype": "int32",
+  "meta": {
+    "quantizer": [
+      "gptqmodel:1.9.0"
+    ],
+    "uri": "https://github.com/modelcloud/gptqmodel",
+    "damp_percent": 0.01,
+    "damp_auto_increment": 0.0025,
+    "static_groups": false,
+    "true_sequential": true,
+    "mse": 0.0
+  }
+}

variant_b/eval/dyk_mc_results.json

@@ -0,0 +1,109 @@
+{
+  "results": {
+    "polish_dyk_multiple_choice": {
+      "acc,none": 0.6287657920310982,
+      "acc_stderr,none": 0.01506856478731849,
+      "f1,none": 0.22672064777327935,
+      "f1_stderr,none": "N/A",
+      "acc_norm,none": 0.6287657920310982,
+      "acc_norm_stderr,none": 0.01506856478731849,
+      "alias": "polish_dyk_multiple_choice"
+    }
+  },
+  "group_subtasks": {
+    "polish_dyk_multiple_choice": []
+  },
+  "configs": {
+    "polish_dyk_multiple_choice": {
+      "task": "polish_dyk_multiple_choice",
+      "dataset_path": "allegro/klej-dyk",
+      "training_split": "train",
+      "test_split": "test",
+      "doc_to_text": "Pytanie: \"{{question}}\"\nSugerowana odpowiedź: \"{{answer}}\"\nPytanie: Czy sugerowana odpowiedź na zadane pytanie jest poprawna?\nOdpowiedz krótko \"Tak\" lub \"Nie\". Prawidłowa odpowiedź:",
+      "doc_to_target": "{{target|int}}",
+      "doc_to_choice": [
+        "Nie",
+        "Tak"
+      ],
+      "description": "",
+      "target_delimiter": " ",
+      "fewshot_delimiter": "\n\n",
+      "num_fewshot": 5,
+      "metric_list": [
+        {
+          "metric": "acc",
+          "aggregation": "mean",
+          "higher_is_better": true
+        },
+        {
+          "metric": "acc_norm",
+          "aggregation": "mean",
+          "higher_is_better": true
+        },
+        {
+          "metric": "def f1(predictions, references):\n    _prediction = predictions[0]\n    _reference = references[0]\n    string_label = [\"B\", \"C\"]\n    reference = string_label.index(_reference)\n    prediction = (\n        string_label.index(_prediction)\n        if _prediction in string_label\n        else 0\n    )\n\n    return (prediction, reference)\n",
+          "aggregation": "def agg_f1(items):\n    predictions, references = zip(*items)\n    references, predictions = np.asarray(references), np.asarray(predictions)\n\n    return sklearn.metrics.f1_score(references, predictions)\n",
+          "higher_is_better": true
+        }
+      ],
+      "output_type": "multiple_choice",
+      "repeats": 1,
+      "should_decontaminate": true,
+      "doc_to_decontamination_query": "{{question}} {{answer}}"
+    }
+  },
+  "versions": {
+    "polish_dyk_multiple_choice": "Yaml"
+  },
+  "n-shot": {
+    "polish_dyk_multiple_choice": 5
+  },
+  "higher_is_better": {
+    "polish_dyk_multiple_choice": {
+      "acc": true,
+      "acc_norm": true,
+      "f1": true
+    }
+  },
+  "n-samples": {
+    "polish_dyk_multiple_choice": {
+      "original": 1029,
+      "effective": 1029
+    }
+  },
+  "config": {
+    "model": "hf",
+    "model_args": "pretrained=/dev/shm/spinquant/exported_model,trust_remote_code=True",
+    "model_num_parameters": 263606272,
+    "model_dtype": "torch.bfloat16",
+    "model_revision": "main",
+    "model_sha": "",
+    "batch_size": "8",
+    "batch_sizes": [],
+    "device": null,
+    "use_cache": null,
+    "limit": null,
+    "bootstrap_iters": 100000,
+    "gen_kwargs": null,
+    "random_seed": 0,
+    "numpy_seed": 1234,
+    "torch_seed": 1234,
+    "fewshot_seed": 1234
+  },
+  "git_hash": "29a34b7",
+  "date": 1771721555.3551857,
+  "pretty_env_info": "PyTorch version: 2.10.0+cu128\nIs debug build: False\nCUDA used to build PyTorch: 12.8\nROCM used to build PyTorch: N/A\n\nOS: Ubuntu 24.04.3 LTS (x86_64)\nGCC version: (Ubuntu 13.3.0-6ubuntu2~24.04) 13.3.0\nClang version: Could not collect\nCMake version: version 3.28.3\nLibc version: glibc-2.39\n\nPython version: 3.12.3 (main, Nov  6 2025, 13:44:16) [GCC 13.3.0] (64-bit runtime)\nPython platform: Linux-6.8.0-90-generic-x86_64-with-glibc2.39\nIs CUDA available: True\nCUDA runtime version: 12.8.93\nCUDA_MODULE_LOADING set to: \nGPU models and configuration: GPU 0: NVIDIA H200\nNvidia driver version: 570.211.01\ncuDNN version: Probably one of the following:\n/usr/lib/x86_64-linux-gnu/libcudnn.so.9.8.0\n/usr/lib/x86_64-linux-gnu/libcudnn_adv.so.9.8.0\n/usr/lib/x86_64-linux-gnu/libcudnn_cnn.so.9.8.0\n/usr/lib/x86_64-linux-gnu/libcudnn_engines_precompiled.so.9.8.0\n/usr/lib/x86_64-linux-gnu/libcudnn_engines_runtime_compiled.so.9.8.0\n/usr/lib/x86_64-linux-gnu/libcudnn_graph.so.9.8.0\n/usr/lib/x86_64-linux-gnu/libcudnn_heuristic.so.9.8.0\n/usr/lib/x86_64-linux-gnu/libcudnn_ops.so.9.8.0\nIs XPU available: False\nHIP runtime version: N/A\nMIOpen runtime version: N/A\nIs XNNPACK available: True\nCaching allocator config: N/A\n\nCPU:\nArchitecture:                         x86_64\nCPU op-mode(s):                       32-bit, 64-bit\nAddress sizes:                        46 bits physical, 57 bits virtual\nByte Order:                           Little Endian\nCPU(s):                               96\nOn-line CPU(s) list:                  0-95\nVendor ID:                            GenuineIntel\nModel name:                           INTEL(R) XEON(R) PLATINUM 8568Y+\nCPU family:                           6\nModel:                                207\nThread(s) per core:                   1\nCore(s) per socket:                   48\nSocket(s):                            2\nStepping:                             2\nCPU(s) scaling MHz:                   35%\nCPU max MHz:                          4000.0000\nCPU min MHz:                          800.0000\nBogoMIPS:                             4600.00\nFlags:                                fpu vme de pse tsc msr pae mce cx8 apic sep mtrr pge mca cmov pat pse36 clflush dts acpi mmx fxsr sse sse2 ss ht tm pbe syscall nx pdpe1gb rdtscp lm constant_tsc art arch_perfmon pebs bts rep_good nopl xtopology nonstop_tsc cpuid aperfmperf tsc_known_freq pni pclmulqdq dtes64 monitor ds_cpl smx est tm2 ssse3 sdbg fma cx16 xtpr pdcm pcid dca sse4_1 sse4_2 x2apic movbe popcnt tsc_deadline_timer aes xsave avx f16c rdrand lahf_lm abm 3dnowprefetch cpuid_fault epb cat_l3 cat_l2 cdp_l3 cdp_l2 ssbd mba ibrs ibpb stibp ibrs_enhanced fsgsbase tsc_adjust bmi1 avx2 smep bmi2 erms invpcid cqm rdt_a avx512f avx512dq rdseed adx smap avx512ifma clflushopt clwb intel_pt avx512cd sha_ni avx512bw avx512vl xsaveopt xsavec xgetbv1 xsaves cqm_llc cqm_occup_llc cqm_mbm_total cqm_mbm_local split_lock_detect user_shstk avx_vnni avx512_bf16 wbnoinvd dtherm ida arat pln pts hfi avx512vbmi umip pku ospke waitpkg avx512_vbmi2 gfni vaes vpclmulqdq avx512_vnni avx512_bitalg tme avx512_vpopcntdq la57 rdpid bus_lock_detect cldemote movdiri movdir64b enqcmd fsrm md_clear serialize tsxldtrk pconfig arch_lbr ibt amx_bf16 avx512_fp16 amx_tile amx_int8 flush_l1d arch_capabilities ibpb_exit_to_user\nL1d cache:                            4.5 MiB (96 instances)\nL1i cache:                            3 MiB (96 instances)\nL2 cache:                             192 MiB (96 instances)\nL3 cache:                             600 MiB (2 instances)\nNUMA node(s):                         4\nNUMA node0 CPU(s):                    0-23\nNUMA node1 CPU(s):                    24-47\nNUMA node2 CPU(s):                    48-71\nNUMA node3 CPU(s):                    72-95\nVulnerability Gather data sampling:   Not affected\nVulnerability Itlb multihit:          Not affected\nVulnerability L1tf:                   Not affected\nVulnerability Mds:                    Not affected\nVulnerability Meltdown:               Not affected\nVulnerability Mmio stale data:        Not affected\nVulnerability Reg file data sampling: Not affected\nVulnerability Retbleed:               Not affected\nVulnerability Spec rstack overflow:   Not affected\nVulnerability Spec store bypass:      Mitigation; Speculative Store Bypass disabled via prctl\nVulnerability Spectre v1:             Mitigation; usercopy/swapgs barriers and __user pointer sanitization\nVulnerability Spectre v2:             Mitigation; Enhanced / Automatic IBRS; IBPB conditional; RSB filling; PBRSB-eIBRS SW sequence; BHI BHI_DIS_S\nVulnerability Srbds:                  Not affected\nVulnerability Tsx async abort:        Not affected\nVulnerability Vmscape:                Mitigation; IBPB before exit to userspace\n\nVersions of relevant libraries:\n[pip3] numpy==1.26.4\n[pip3] nvidia-cublas-cu12==12.8.4.1\n[pip3] nvidia-cuda-cupti-cu12==12.8.90\n[pip3] nvidia-cuda-nvrtc-cu12==12.8.93\n[pip3] nvidia-cuda-runtime-cu12==12.8.90\n[pip3] nvidia-cudnn-cu12==9.10.2.21\n[pip3] nvidia-cufft-cu12==11.3.3.83\n[pip3] nvidia-curand-cu12==10.3.9.90\n[pip3] nvidia-cusolver-cu12==11.7.3.90\n[pip3] nvidia-cusparse-cu12==12.5.8.93\n[pip3] nvidia-cusparselt-cu12==0.7.1\n[pip3] nvidia-nccl-cu12==2.27.5\n[pip3] nvidia-nvjitlink-cu12==12.8.93\n[pip3] nvidia-nvtx-cu12==12.8.90\n[pip3] torch==2.10.0+cu128\n[pip3] torchaudio==2.6.0+cu124\n[pip3] torchvision==0.25.0+cu128\n[pip3] triton==3.6.0\n[conda] Could not collect",
+  "transformers_version": "4.43.4",
+  "upper_git_hash": null,
+  "task_hashes": {},
+  "model_source": "hf",
+  "model_name": "/dev/shm/spinquant/exported_model",
+  "model_name_sanitized": "__dev__shm__spinquant__exported_model",
+  "system_instruction": null,
+  "system_instruction_sha": null,
+  "chat_template": null,
+  "chat_template_sha": null,
+  "start_time": 2027078.064227919,
+  "end_time": 2027209.734849108,
+  "total_evaluation_time_seconds": "131.67062118882313"
+}

variant_b/logs/full_eval.log

@@ -0,0 +1,16 @@
+=== POLISH_MC START Sun Feb 22 00:57:26 UTC 2026 ===
+2026-02-22:00:57:31,125 INFO     [__main__.py:272] Verbosity set to INFO
+2026-02-22:00:57:34,125 INFO     [__main__.py:363] Selected Tasks: ['polish_mc']
+2026-02-22:00:57:34,127 INFO     [evaluator.py:152] Setting random seed to 0 | Setting numpy seed to 1234 | Setting torch manual seed to 1234
+2026-02-22:00:57:34,127 INFO     [evaluator.py:189] Initializing hf model, with arguments: {'pretrained': '/dev/shm/spinquant/exported_model', 'trust_remote_code': True}
+2026-02-22:00:57:34,489 INFO     [huggingface.py:169] Using device 'cuda'
+/usr/local/lib/python3.12/dist-packages/auto_gptq/nn_modules/triton_utils/kernels.py:410: FutureWarning: `torch.cuda.amp.custom_fwd(args...)` is deprecated. Please use `torch.amp.custom_fwd(args..., device_type='cuda')` instead.
+  @custom_fwd
+/usr/local/lib/python3.12/dist-packages/auto_gptq/nn_modules/triton_utils/kernels.py:418: FutureWarning: `torch.cuda.amp.custom_bwd(args...)` is deprecated. Please use `torch.amp.custom_bwd(args..., device_type='cuda')` instead.
+  @custom_bwd
+/usr/local/lib/python3.12/dist-packages/auto_gptq/nn_modules/triton_utils/kernels.py:461: FutureWarning: `torch.cuda.amp.custom_fwd(args...)` is deprecated. Please use `torch.amp.custom_fwd(args..., device_type='cuda')` instead.
+  @custom_fwd(cast_inputs=torch.float16)
+2026-02-22:00:57:34,742 WARNING  [qlinear_cuda.py:18] CUDA extension not installed.
+2026-02-22:00:57:34,742 WARNING  [qlinear_cuda_old.py:17] CUDA extension not installed.
+/usr/local/lib/python3.12/dist-packages/transformers/modeling_utils.py:4674: FutureWarning: `_is_quantized_training_enabled` is going to be deprecated in transformers 4.39.0. Please use `model.hf_quantizer.is_trainable` instead
+  warnings.warn(

variant_b/logs/polish_mc.log

@@ -0,0 +1,63 @@
+2026-02-22:00:55:04,271 INFO     [__main__.py:272] Verbosity set to INFO
+2026-02-22:00:55:07,285 INFO     [__main__.py:363] Selected Tasks: ['polish_mc']
+2026-02-22:00:55:07,287 INFO     [evaluator.py:152] Setting random seed to 0 | Setting numpy seed to 1234 | Setting torch manual seed to 1234
+2026-02-22:00:55:07,287 INFO     [evaluator.py:189] Initializing hf model, with arguments: {'pretrained': '/dev/shm/spinquant/exported_model', 'trust_remote_code': True}
+2026-02-22:00:55:07,649 INFO     [huggingface.py:169] Using device 'cuda'
+/usr/local/lib/python3.12/dist-packages/auto_gptq/nn_modules/triton_utils/kernels.py:410: FutureWarning: `torch.cuda.amp.custom_fwd(args...)` is deprecated. Please use `torch.amp.custom_fwd(args..., device_type='cuda')` instead.
+  @custom_fwd
+/usr/local/lib/python3.12/dist-packages/auto_gptq/nn_modules/triton_utils/kernels.py:418: FutureWarning: `torch.cuda.amp.custom_bwd(args...)` is deprecated. Please use `torch.amp.custom_bwd(args..., device_type='cuda')` instead.
+  @custom_bwd
+/usr/local/lib/python3.12/dist-packages/auto_gptq/nn_modules/triton_utils/kernels.py:461: FutureWarning: `torch.cuda.amp.custom_fwd(args...)` is deprecated. Please use `torch.amp.custom_fwd(args..., device_type='cuda')` instead.
+  @custom_fwd(cast_inputs=torch.float16)
+2026-02-22:00:55:07,892 WARNING  [qlinear_cuda.py:18] CUDA extension not installed.
+2026-02-22:00:55:07,893 WARNING  [qlinear_cuda_old.py:17] CUDA extension not installed.
+/usr/local/lib/python3.12/dist-packages/transformers/modeling_utils.py:4674: FutureWarning: `_is_quantized_training_enabled` is going to be deprecated in transformers 4.39.0. Please use `model.hf_quantizer.is_trainable` instead
+  warnings.warn(
+Traceback (most recent call last):
+  File "<frozen runpy>", line 198, in _run_module_as_main
+  File "<frozen runpy>", line 88, in _run_code
+  File "/dev/shm/eval/lm-evaluation-harness/lm_eval/__main__.py", line 448, in <module>
+    cli_evaluate()
+  File "/dev/shm/eval/lm-evaluation-harness/lm_eval/__main__.py", line 369, in cli_evaluate
+    results = evaluator.simple_evaluate(
+              ^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/dev/shm/eval/lm-evaluation-harness/lm_eval/utils.py", line 346, in _wrapper
+    return fn(*args, **kwargs)
+           ^^^^^^^^^^^^^^^^^^^
+  File "/dev/shm/eval/lm-evaluation-harness/lm_eval/evaluator.py", line 221, in simple_evaluate
+    task_dict = get_task_dict(tasks, task_manager)
+                ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/dev/shm/eval/lm-evaluation-harness/lm_eval/tasks/__init__.py", line 423, in get_task_dict
+    task_name_from_string_dict = task_manager.load_task_or_group(
+                                 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/dev/shm/eval/lm-evaluation-harness/lm_eval/tasks/__init__.py", line 271, in load_task_or_group
+    collections.ChainMap(*map(self._load_individual_task_or_group, task_list))
+  File "/dev/shm/eval/lm-evaluation-harness/lm_eval/tasks/__init__.py", line 254, in _load_individual_task_or_group
+    **dict(collections.ChainMap(*map(fn, subtask_list))),
+           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/dev/shm/eval/lm-evaluation-harness/lm_eval/tasks/__init__.py", line 162, in _load_individual_task_or_group
+    return load_task(task_config, task=name_or_config, group=parent_name)
+           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/dev/shm/eval/lm-evaluation-harness/lm_eval/tasks/__init__.py", line 151, in load_task
+    task_object = ConfigurableTask(config=config)
+                  ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/dev/shm/eval/lm-evaluation-harness/lm_eval/api/task.py", line 809, in __init__
+    self.download(self.config.dataset_kwargs)
+  File "/dev/shm/eval/lm-evaluation-harness/lm_eval/api/task.py", line 916, in download
+    self.dataset = datasets.load_dataset(
+                   ^^^^^^^^^^^^^^^^^^^^^^
+  File "/usr/local/lib/python3.12/dist-packages/datasets/load.py", line 1488, in load_dataset
+    builder_instance = load_dataset_builder(
+                       ^^^^^^^^^^^^^^^^^^^^^
+  File "/usr/local/lib/python3.12/dist-packages/datasets/load.py", line 1167, in load_dataset_builder
+    builder_instance: DatasetBuilder = builder_cls(
+                                       ^^^^^^^^^^^^
+  File "/usr/local/lib/python3.12/dist-packages/datasets/builder.py", line 343, in __init__
+    self.config, self.config_id = self._create_builder_config(
+                                  ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/usr/local/lib/python3.12/dist-packages/datasets/builder.py", line 515, in _create_builder_config
+    raise ValueError(
+ValueError: Config name is missing.
+Please pick one among the available configs: ['acm_Arab', 'arz_Arab', 'ceb_Latn', 'fin_Latn', 'hin_Deva', 'ita_Latn', 'khm_Khmr', 'lvs_Latn', 'npi_Deva', 'pol_Latn', 'slv_Latn', 'swe_Latn', 'tso_Latn', 'xho_Latn', 'afr_Latn', 'asm_Beng', 'ces_Latn', 'fra_Latn', 'hin_Latn', 'jav_Latn', 'kin_Latn', 'mal_Mlym', 'npi_Latn', 'por_Latn', 'sna_Latn', 'swh_Latn', 'tur_Latn', 'yor_Latn', 'als_Latn', 'azj_Latn', 'ckb_Arab', 'fuv_Latn', 'hrv_Latn', 'jpn_Jpan', 'kir_Cyrl', 'mar_Deva', 'nso_Latn', 'snd_Arab', 'tam_Taml', 'ukr_Cyrl', 'zho_Hans', 'amh_Ethi', 'bam_Latn', 'dan_Latn', 'gaz_Latn', 'hun_Latn', 'kac_Latn', 'kor_Hang', 'mkd_Cyrl', 'nya_Latn', 'ron_Latn', 'som_Latn', 'tel_Telu', 'urd_Arab', 'zho_Hant', 'apc_Arab', 'ben_Beng', 'deu_Latn', 'grn_Latn', 'hye_Armn', 'kan_Knda', 'lao_Laoo', 'mlt_Latn', 'ory_Orya', 'rus_Cyrl', 'sot_Latn', 'tgk_Cyrl', 'urd_Latn', 'zsm_Latn', 'arb_Arab', 'ben_Latn', 'ell_Grek', 'guj_Gujr', 'ibo_Latn', 'kat_Geor', 'lin_Latn', 'mri_Latn', 'pan_Guru', 'shn_Mymr', 'spa_Latn', 'tgl_Latn', 'uzn_Latn', 'zul_Latn', 'arb_Latn', 'bod_Tibt', 'eng_Latn', 'hat_Latn', 'ilo_Latn', 'kaz_Cyrl', 'lit_Latn', 'mya_Mymr', 'pbt_Arab', 'sin_Latn', 'srp_Cyrl', 'tha_Thai', 'vie_Latn', 'ars_Arab', 'bul_Cyrl', 'est_Latn', 'hau_Latn', 'ind_Latn', 'kea_Latn', 'lug_Latn', 'nld_Latn', 'pes_Arab', 'sin_Sinh', 'ssw_Latn', 'tir_Ethi', 'war_Latn', 'ary_Arab', 'cat_Latn', 'eus_Latn', 'heb_Hebr', 'isl_Latn', 'khk_Cyrl', 'luo_Latn', 'nob_Latn', 'plt_Latn', 'slk_Latn', 'sun_Latn', 'tsn_Latn', 'wol_Latn']
+Example of usage:
+	`load_dataset('facebook/belebele', 'acm_Arab')`

variant_b/logs/step4b_output.log

@@ -0,0 +1,468 @@
+[2026-02-21 20:06:46] Step 4b RESTARTED: AutoGPTQ 2-bit quantization (fixed buffer device)...
+/usr/local/lib/python3.12/dist-packages/auto_gptq/nn_modules/triton_utils/kernels.py:410: FutureWarning: `torch.cuda.amp.custom_fwd(args...)` is deprecated. Please use `torch.amp.custom_fwd(args..., device_type='cuda')` instead.
+  @custom_fwd
+/usr/local/lib/python3.12/dist-packages/auto_gptq/nn_modules/triton_utils/kernels.py:418: FutureWarning: `torch.cuda.amp.custom_bwd(args...)` is deprecated. Please use `torch.amp.custom_bwd(args..., device_type='cuda')` instead.
+  @custom_bwd
+/usr/local/lib/python3.12/dist-packages/auto_gptq/nn_modules/triton_utils/kernels.py:461: FutureWarning: `torch.cuda.amp.custom_fwd(args...)` is deprecated. Please use `torch.amp.custom_fwd(args..., device_type='cuda')` instead.
+  @custom_fwd(cast_inputs=torch.float16)
+CUDA extension not installed.
+CUDA extension not installed.
+Loading calibration data...
+Calibration: 128 samples
+GPTQ config: bits=2, group_size=128, desc_act=True, sym=True
+Loading rotated model for GPTQ...
+
+Moving all model buffers to CUDA...
+  Moving buffer model.layers.0.self_attn.rotary_emb.inv_freq from CPU to CUDA
+  Moving buffer model.layers.1.self_attn.rotary_emb.inv_freq from CPU to CUDA
+  Moving buffer model.layers.2.self_attn.rotary_emb.inv_freq from CPU to CUDA
+  Moving buffer model.layers.3.self_attn.rotary_emb.inv_freq from CPU to CUDA
+  Moving buffer model.layers.4.self_attn.rotary_emb.inv_freq from CPU to CUDA
+  Moving buffer model.layers.5.self_attn.rotary_emb.inv_freq from CPU to CUDA
+  Moving buffer model.layers.6.self_attn.rotary_emb.inv_freq from CPU to CUDA
+  Moving buffer model.layers.7.self_attn.rotary_emb.inv_freq from CPU to CUDA
+  Moving buffer model.layers.8.self_attn.rotary_emb.inv_freq from CPU to CUDA
+  Moving buffer model.layers.9.self_attn.rotary_emb.inv_freq from CPU to CUDA
+  Moving buffer model.layers.10.self_attn.rotary_emb.inv_freq from CPU to CUDA
+  Moving buffer model.layers.11.self_attn.rotary_emb.inv_freq from CPU to CUDA
+  Moving buffer model.layers.12.self_attn.rotary_emb.inv_freq from CPU to CUDA
+  Moving buffer model.layers.13.self_attn.rotary_emb.inv_freq from CPU to CUDA
+  Moving buffer model.layers.14.self_attn.rotary_emb.inv_freq from CPU to CUDA
+  Moving buffer model.layers.15.self_attn.rotary_emb.inv_freq from CPU to CUDA
+  Moving buffer model.layers.16.self_attn.rotary_emb.inv_freq from CPU to CUDA
+  Moving buffer model.layers.17.self_attn.rotary_emb.inv_freq from CPU to CUDA
+  Moving buffer model.layers.18.self_attn.rotary_emb.inv_freq from CPU to CUDA
+  Moving buffer model.layers.19.self_attn.rotary_emb.inv_freq from CPU to CUDA
+  Moving buffer model.layers.20.self_attn.rotary_emb.inv_freq from CPU to CUDA
+  Moving buffer model.layers.21.self_attn.rotary_emb.inv_freq from CPU to CUDA
+  Moving buffer model.layers.22.self_attn.rotary_emb.inv_freq from CPU to CUDA
+  Moving buffer model.layers.23.self_attn.rotary_emb.inv_freq from CPU to CUDA
+  Moving buffer model.layers.24.self_attn.rotary_emb.inv_freq from CPU to CUDA
+  Moving buffer model.layers.25.self_attn.rotary_emb.inv_freq from CPU to CUDA
+  Moving buffer model.layers.26.self_attn.rotary_emb.inv_freq from CPU to CUDA
+  Moving buffer model.layers.27.self_attn.rotary_emb.inv_freq from CPU to CUDA
+  Moving buffer model.layers.28.self_attn.rotary_emb.inv_freq from CPU to CUDA
+  Moving buffer model.layers.29.self_attn.rotary_emb.inv_freq from CPU to CUDA
+  Moving buffer model.layers.30.self_attn.rotary_emb.inv_freq from CPU to CUDA
+  Moving buffer model.layers.31.self_attn.rotary_emb.inv_freq from CPU to CUDA
+  Moving buffer model.layers.32.self_attn.rotary_emb.inv_freq from CPU to CUDA
+  Moving buffer model.layers.33.self_attn.rotary_emb.inv_freq from CPU to CUDA
+  Moving buffer model.layers.34.self_attn.rotary_emb.inv_freq from CPU to CUDA
+  Moving buffer model.layers.35.self_attn.rotary_emb.inv_freq from CPU to CUDA
+  Moving buffer model.layers.36.self_attn.rotary_emb.inv_freq from CPU to CUDA
+  Moving buffer model.layers.37.self_attn.rotary_emb.inv_freq from CPU to CUDA
+  Moving buffer model.layers.38.self_attn.rotary_emb.inv_freq from CPU to CUDA
+  Moving buffer model.layers.39.self_attn.rotary_emb.inv_freq from CPU to CUDA
+  Moving buffer model.layers.40.self_attn.rotary_emb.inv_freq from CPU to CUDA
+  Moving buffer model.layers.41.self_attn.rotary_emb.inv_freq from CPU to CUDA
+  Moving buffer model.layers.42.self_attn.rotary_emb.inv_freq from CPU to CUDA
+  Moving buffer model.layers.43.self_attn.rotary_emb.inv_freq from CPU to CUDA
+  Moving buffer model.layers.44.self_attn.rotary_emb.inv_freq from CPU to CUDA
+  Moving buffer model.layers.45.self_attn.rotary_emb.inv_freq from CPU to CUDA
+  Moving buffer model.layers.46.self_attn.rotary_emb.inv_freq from CPU to CUDA
+  Moving buffer model.layers.47.self_attn.rotary_emb.inv_freq from CPU to CUDA
+  Moving buffer model.layers.48.self_attn.rotary_emb.inv_freq from CPU to CUDA
+  Moving buffer model.layers.49.self_attn.rotary_emb.inv_freq from CPU to CUDA
+  Moving buffer model.rotary_emb.inv_freq from CPU to CUDA
+Running GPTQ 2-bit quantization...
+INFO - Start quantizing layer 1/50
+INFO - Quantizing self_attn.k_proj in layer 1/50...
+INFO - Quantizing self_attn.v_proj in layer 1/50...
+INFO - Quantizing self_attn.q_proj in layer 1/50...
+INFO - Quantizing self_attn.o_proj in layer 1/50...
+INFO - Quantizing mlp.up_proj in layer 1/50...
+INFO - Quantizing mlp.gate_proj in layer 1/50...
+INFO - Quantizing mlp.down_proj in layer 1/50...
+INFO - Start quantizing layer 2/50
+INFO - Quantizing self_attn.k_proj in layer 2/50...
+INFO - Quantizing self_attn.v_proj in layer 2/50...
+INFO - Quantizing self_attn.q_proj in layer 2/50...
+INFO - Quantizing self_attn.o_proj in layer 2/50...
+INFO - Quantizing mlp.up_proj in layer 2/50...
+INFO - Quantizing mlp.gate_proj in layer 2/50...
+INFO - Quantizing mlp.down_proj in layer 2/50...
+INFO - Start quantizing layer 3/50
+INFO - Quantizing self_attn.k_proj in layer 3/50...
+INFO - Quantizing self_attn.v_proj in layer 3/50...
+INFO - Quantizing self_attn.q_proj in layer 3/50...
+INFO - Quantizing self_attn.o_proj in layer 3/50...
+INFO - Quantizing mlp.up_proj in layer 3/50...
+INFO - Quantizing mlp.gate_proj in layer 3/50...
+INFO - Quantizing mlp.down_proj in layer 3/50...
+INFO - Start quantizing layer 4/50
+INFO - Quantizing self_attn.k_proj in layer 4/50...
+INFO - Quantizing self_attn.v_proj in layer 4/50...
+INFO - Quantizing self_attn.q_proj in layer 4/50...
+INFO - Quantizing self_attn.o_proj in layer 4/50...
+INFO - Quantizing mlp.up_proj in layer 4/50...
+INFO - Quantizing mlp.gate_proj in layer 4/50...
+INFO - Quantizing mlp.down_proj in layer 4/50...
+INFO - Start quantizing layer 5/50
+INFO - Quantizing self_attn.k_proj in layer 5/50...
+INFO - Quantizing self_attn.v_proj in layer 5/50...
+INFO - Quantizing self_attn.q_proj in layer 5/50...
+INFO - Quantizing self_attn.o_proj in layer 5/50...
+INFO - Quantizing mlp.up_proj in layer 5/50...
+INFO - Quantizing mlp.gate_proj in layer 5/50...
+INFO - Quantizing mlp.down_proj in layer 5/50...
+INFO - Start quantizing layer 6/50
+INFO - Quantizing self_attn.k_proj in layer 6/50...
+INFO - Quantizing self_attn.v_proj in layer 6/50...
+INFO - Quantizing self_attn.q_proj in layer 6/50...
+INFO - Quantizing self_attn.o_proj in layer 6/50...
+INFO - Quantizing mlp.up_proj in layer 6/50...
+INFO - Quantizing mlp.gate_proj in layer 6/50...
+INFO - Quantizing mlp.down_proj in layer 6/50...
+INFO - Start quantizing layer 7/50
+INFO - Quantizing self_attn.k_proj in layer 7/50...
+INFO - Quantizing self_attn.v_proj in layer 7/50...
+INFO - Quantizing self_attn.q_proj in layer 7/50...
+INFO - Quantizing self_attn.o_proj in layer 7/50...
+INFO - Quantizing mlp.up_proj in layer 7/50...
+INFO - Quantizing mlp.gate_proj in layer 7/50...
+INFO - Quantizing mlp.down_proj in layer 7/50...
+INFO - Start quantizing layer 8/50
+INFO - Quantizing self_attn.k_proj in layer 8/50...
+INFO - Quantizing self_attn.v_proj in layer 8/50...
+INFO - Quantizing self_attn.q_proj in layer 8/50...
+INFO - Quantizing self_attn.o_proj in layer 8/50...
+INFO - Quantizing mlp.up_proj in layer 8/50...
+INFO - Quantizing mlp.gate_proj in layer 8/50...
+INFO - Quantizing mlp.down_proj in layer 8/50...
+INFO - Start quantizing layer 9/50
+INFO - Quantizing self_attn.k_proj in layer 9/50...
+INFO - Quantizing self_attn.v_proj in layer 9/50...
+INFO - Quantizing self_attn.q_proj in layer 9/50...
+INFO - Quantizing self_attn.o_proj in layer 9/50...
+INFO - Quantizing mlp.up_proj in layer 9/50...
+INFO - Quantizing mlp.gate_proj in layer 9/50...
+INFO - Quantizing mlp.down_proj in layer 9/50...
+INFO - Start quantizing layer 10/50
+INFO - Quantizing self_attn.k_proj in layer 10/50...
+INFO - Quantizing self_attn.v_proj in layer 10/50...
+INFO - Quantizing self_attn.q_proj in layer 10/50...
+INFO - Quantizing self_attn.o_proj in layer 10/50...
+INFO - Quantizing mlp.up_proj in layer 10/50...
+INFO - Quantizing mlp.gate_proj in layer 10/50...
+INFO - Quantizing mlp.down_proj in layer 10/50...
+INFO - Start quantizing layer 11/50
+INFO - Quantizing self_attn.k_proj in layer 11/50...
+INFO - Quantizing self_attn.v_proj in layer 11/50...
+INFO - Quantizing self_attn.q_proj in layer 11/50...
+INFO - Quantizing self_attn.o_proj in layer 11/50...
+INFO - Quantizing mlp.up_proj in layer 11/50...
+INFO - Quantizing mlp.gate_proj in layer 11/50...
+INFO - Quantizing mlp.down_proj in layer 11/50...
+INFO - Start quantizing layer 12/50
+INFO - Quantizing self_attn.k_proj in layer 12/50...
+INFO - Quantizing self_attn.v_proj in layer 12/50...
+INFO - Quantizing self_attn.q_proj in layer 12/50...
+INFO - Quantizing self_attn.o_proj in layer 12/50...
+INFO - Quantizing mlp.up_proj in layer 12/50...
+INFO - Quantizing mlp.gate_proj in layer 12/50...
+INFO - Quantizing mlp.down_proj in layer 12/50...
+INFO - Start quantizing layer 13/50
+INFO - Quantizing self_attn.k_proj in layer 13/50...
+INFO - Quantizing self_attn.v_proj in layer 13/50...
+INFO - Quantizing self_attn.q_proj in layer 13/50...
+INFO - Quantizing self_attn.o_proj in layer 13/50...
+INFO - Quantizing mlp.up_proj in layer 13/50...
+INFO - Quantizing mlp.gate_proj in layer 13/50...
+INFO - Quantizing mlp.down_proj in layer 13/50...
+INFO - Start quantizing layer 14/50
+INFO - Quantizing self_attn.k_proj in layer 14/50...
+INFO - Quantizing self_attn.v_proj in layer 14/50...
+INFO - Quantizing self_attn.q_proj in layer 14/50...
+INFO - Quantizing self_attn.o_proj in layer 14/50...
+INFO - Quantizing mlp.up_proj in layer 14/50...
+INFO - Quantizing mlp.gate_proj in layer 14/50...
+INFO - Quantizing mlp.down_proj in layer 14/50...
+INFO - Start quantizing layer 15/50
+INFO - Quantizing self_attn.k_proj in layer 15/50...
+INFO - Quantizing self_attn.v_proj in layer 15/50...
+INFO - Quantizing self_attn.q_proj in layer 15/50...
+INFO - Quantizing self_attn.o_proj in layer 15/50...
+INFO - Quantizing mlp.up_proj in layer 15/50...
+INFO - Quantizing mlp.gate_proj in layer 15/50...
+INFO - Quantizing mlp.down_proj in layer 15/50...
+INFO - Start quantizing layer 16/50
+INFO - Quantizing self_attn.k_proj in layer 16/50...
+INFO - Quantizing self_attn.v_proj in layer 16/50...
+INFO - Quantizing self_attn.q_proj in layer 16/50...
+INFO - Quantizing self_attn.o_proj in layer 16/50...
+INFO - Quantizing mlp.up_proj in layer 16/50...
+INFO - Quantizing mlp.gate_proj in layer 16/50...
+INFO - Quantizing mlp.down_proj in layer 16/50...
+INFO - Start quantizing layer 17/50
+INFO - Quantizing self_attn.k_proj in layer 17/50...
+INFO - Quantizing self_attn.v_proj in layer 17/50...
+INFO - Quantizing self_attn.q_proj in layer 17/50...
+INFO - Quantizing self_attn.o_proj in layer 17/50...
+INFO - Quantizing mlp.up_proj in layer 17/50...
+INFO - Quantizing mlp.gate_proj in layer 17/50...
+INFO - Quantizing mlp.down_proj in layer 17/50...
+INFO - Start quantizing layer 18/50
+INFO - Quantizing self_attn.k_proj in layer 18/50...
+INFO - Quantizing self_attn.v_proj in layer 18/50...
+INFO - Quantizing self_attn.q_proj in layer 18/50...
+INFO - Quantizing self_attn.o_proj in layer 18/50...
+INFO - Quantizing mlp.up_proj in layer 18/50...
+INFO - Quantizing mlp.gate_proj in layer 18/50...
+INFO - Quantizing mlp.down_proj in layer 18/50...
+INFO - Start quantizing layer 19/50
+INFO - Quantizing self_attn.k_proj in layer 19/50...
+INFO - Quantizing self_attn.v_proj in layer 19/50...
+INFO - Quantizing self_attn.q_proj in layer 19/50...
+INFO - Quantizing self_attn.o_proj in layer 19/50...
+INFO - Quantizing mlp.up_proj in layer 19/50...
+INFO - Quantizing mlp.gate_proj in layer 19/50...
+INFO - Quantizing mlp.down_proj in layer 19/50...
+INFO - Start quantizing layer 20/50
+INFO - Quantizing self_attn.k_proj in layer 20/50...
+INFO - Quantizing self_attn.v_proj in layer 20/50...
+INFO - Quantizing self_attn.q_proj in layer 20/50...
+INFO - Quantizing self_attn.o_proj in layer 20/50...
+INFO - Quantizing mlp.up_proj in layer 20/50...
+INFO - Quantizing mlp.gate_proj in layer 20/50...
+INFO - Quantizing mlp.down_proj in layer 20/50...
+INFO - Start quantizing layer 21/50
+INFO - Quantizing self_attn.k_proj in layer 21/50...
+INFO - Quantizing self_attn.v_proj in layer 21/50...
+INFO - Quantizing self_attn.q_proj in layer 21/50...
+INFO - Quantizing self_attn.o_proj in layer 21/50...
+INFO - Quantizing mlp.up_proj in layer 21/50...
+INFO - Quantizing mlp.gate_proj in layer 21/50...
+INFO - Quantizing mlp.down_proj in layer 21/50...
+INFO - Start quantizing layer 22/50
+INFO - Quantizing self_attn.k_proj in layer 22/50...
+INFO - Quantizing self_attn.v_proj in layer 22/50...
+INFO - Quantizing self_attn.q_proj in layer 22/50...
+INFO - Quantizing self_attn.o_proj in layer 22/50...
+INFO - Quantizing mlp.up_proj in layer 22/50...
+INFO - Quantizing mlp.gate_proj in layer 22/50...
+INFO - Quantizing mlp.down_proj in layer 22/50...
+INFO - Start quantizing layer 23/50
+INFO - Quantizing self_attn.k_proj in layer 23/50...
+INFO - Quantizing self_attn.v_proj in layer 23/50...
+INFO - Quantizing self_attn.q_proj in layer 23/50...
+INFO - Quantizing self_attn.o_proj in layer 23/50...
+INFO - Quantizing mlp.up_proj in layer 23/50...
+INFO - Quantizing mlp.gate_proj in layer 23/50...
+INFO - Quantizing mlp.down_proj in layer 23/50...
+INFO - Start quantizing layer 24/50
+INFO - Quantizing self_attn.k_proj in layer 24/50...
+INFO - Quantizing self_attn.v_proj in layer 24/50...
+INFO - Quantizing self_attn.q_proj in layer 24/50...
+INFO - Quantizing self_attn.o_proj in layer 24/50...
+INFO - Quantizing mlp.up_proj in layer 24/50...
+INFO - Quantizing mlp.gate_proj in layer 24/50...
+INFO - Quantizing mlp.down_proj in layer 24/50...
+INFO - Start quantizing layer 25/50
+INFO - Quantizing self_attn.k_proj in layer 25/50...
+INFO - Quantizing self_attn.v_proj in layer 25/50...
+INFO - Quantizing self_attn.q_proj in layer 25/50...
+INFO - Quantizing self_attn.o_proj in layer 25/50...
+INFO - Quantizing mlp.up_proj in layer 25/50...
+INFO - Quantizing mlp.gate_proj in layer 25/50...
+INFO - Quantizing mlp.down_proj in layer 25/50...
+INFO - Start quantizing layer 26/50
+INFO - Quantizing self_attn.k_proj in layer 26/50...
+INFO - Quantizing self_attn.v_proj in layer 26/50...
+INFO - Quantizing self_attn.q_proj in layer 26/50...
+INFO - Quantizing self_attn.o_proj in layer 26/50...
+INFO - Quantizing mlp.up_proj in layer 26/50...
+INFO - Quantizing mlp.gate_proj in layer 26/50...
+INFO - Quantizing mlp.down_proj in layer 26/50...
+INFO - Start quantizing layer 27/50
+INFO - Quantizing self_attn.k_proj in layer 27/50...
+INFO - Quantizing self_attn.v_proj in layer 27/50...
+INFO - Quantizing self_attn.q_proj in layer 27/50...
+INFO - Quantizing self_attn.o_proj in layer 27/50...
+INFO - Quantizing mlp.up_proj in layer 27/50...
+INFO - Quantizing mlp.gate_proj in layer 27/50...
+INFO - Quantizing mlp.down_proj in layer 27/50...
+INFO - Start quantizing layer 28/50
+INFO - Quantizing self_attn.k_proj in layer 28/50...
+INFO - Quantizing self_attn.v_proj in layer 28/50...
+INFO - Quantizing self_attn.q_proj in layer 28/50...
+INFO - Quantizing self_attn.o_proj in layer 28/50...
+INFO - Quantizing mlp.up_proj in layer 28/50...
+INFO - Quantizing mlp.gate_proj in layer 28/50...
+INFO - Quantizing mlp.down_proj in layer 28/50...
+INFO - Start quantizing layer 29/50
+INFO - Quantizing self_attn.k_proj in layer 29/50...
+INFO - Quantizing self_attn.v_proj in layer 29/50...
+INFO - Quantizing self_attn.q_proj in layer 29/50...
+INFO - Quantizing self_attn.o_proj in layer 29/50...
+INFO - Quantizing mlp.up_proj in layer 29/50...
+INFO - Quantizing mlp.gate_proj in layer 29/50...
+INFO - Quantizing mlp.down_proj in layer 29/50...
+INFO - Start quantizing layer 30/50
+INFO - Quantizing self_attn.k_proj in layer 30/50...
+INFO - Quantizing self_attn.v_proj in layer 30/50...
+INFO - Quantizing self_attn.q_proj in layer 30/50...
+INFO - Quantizing self_attn.o_proj in layer 30/50...
+INFO - Quantizing mlp.up_proj in layer 30/50...
+INFO - Quantizing mlp.gate_proj in layer 30/50...
+INFO - Quantizing mlp.down_proj in layer 30/50...
+INFO - Start quantizing layer 31/50
+INFO - Quantizing self_attn.k_proj in layer 31/50...
+INFO - Quantizing self_attn.v_proj in layer 31/50...
+INFO - Quantizing self_attn.q_proj in layer 31/50...
+INFO - Quantizing self_attn.o_proj in layer 31/50...
+INFO - Quantizing mlp.up_proj in layer 31/50...
+INFO - Quantizing mlp.gate_proj in layer 31/50...
+INFO - Quantizing mlp.down_proj in layer 31/50...
+INFO - Start quantizing layer 32/50
+INFO - Quantizing self_attn.k_proj in layer 32/50...
+INFO - Quantizing self_attn.v_proj in layer 32/50...
+INFO - Quantizing self_attn.q_proj in layer 32/50...
+INFO - Quantizing self_attn.o_proj in layer 32/50...
+INFO - Quantizing mlp.up_proj in layer 32/50...
+INFO - Quantizing mlp.gate_proj in layer 32/50...
+INFO - Quantizing mlp.down_proj in layer 32/50...
+INFO - Start quantizing layer 33/50
+INFO - Quantizing self_attn.k_proj in layer 33/50...
+INFO - Quantizing self_attn.v_proj in layer 33/50...
+INFO - Quantizing self_attn.q_proj in layer 33/50...
+INFO - Quantizing self_attn.o_proj in layer 33/50...
+INFO - Quantizing mlp.up_proj in layer 33/50...
+INFO - Quantizing mlp.gate_proj in layer 33/50...
+INFO - Quantizing mlp.down_proj in layer 33/50...
+INFO - Start quantizing layer 34/50
+INFO - Quantizing self_attn.k_proj in layer 34/50...
+INFO - Quantizing self_attn.v_proj in layer 34/50...
+INFO - Quantizing self_attn.q_proj in layer 34/50...
+INFO - Quantizing self_attn.o_proj in layer 34/50...
+INFO - Quantizing mlp.up_proj in layer 34/50...
+INFO - Quantizing mlp.gate_proj in layer 34/50...
+INFO - Quantizing mlp.down_proj in layer 34/50...
+INFO - Start quantizing layer 35/50
+INFO - Quantizing self_attn.k_proj in layer 35/50...
+INFO - Quantizing self_attn.v_proj in layer 35/50...
+INFO - Quantizing self_attn.q_proj in layer 35/50...
+INFO - Quantizing self_attn.o_proj in layer 35/50...
+INFO - Quantizing mlp.up_proj in layer 35/50...
+INFO - Quantizing mlp.gate_proj in layer 35/50...
+INFO - Quantizing mlp.down_proj in layer 35/50...
+INFO - Start quantizing layer 36/50
+INFO - Quantizing self_attn.k_proj in layer 36/50...
+INFO - Quantizing self_attn.v_proj in layer 36/50...
+INFO - Quantizing self_attn.q_proj in layer 36/50...
+INFO - Quantizing self_attn.o_proj in layer 36/50...
+INFO - Quantizing mlp.up_proj in layer 36/50...
+INFO - Quantizing mlp.gate_proj in layer 36/50...
+INFO - Quantizing mlp.down_proj in layer 36/50...
+INFO - Start quantizing layer 37/50
+INFO - Quantizing self_attn.k_proj in layer 37/50...
+INFO - Quantizing self_attn.v_proj in layer 37/50...
+INFO - Quantizing self_attn.q_proj in layer 37/50...
+INFO - Quantizing self_attn.o_proj in layer 37/50...
+INFO - Quantizing mlp.up_proj in layer 37/50...
+INFO - Quantizing mlp.gate_proj in layer 37/50...
+INFO - Quantizing mlp.down_proj in layer 37/50...
+INFO - Start quantizing layer 38/50
+INFO - Quantizing self_attn.k_proj in layer 38/50...
+INFO - Quantizing self_attn.v_proj in layer 38/50...
+INFO - Quantizing self_attn.q_proj in layer 38/50...
+INFO - Quantizing self_attn.o_proj in layer 38/50...
+INFO - Quantizing mlp.up_proj in layer 38/50...
+INFO - Quantizing mlp.gate_proj in layer 38/50...
+INFO - Quantizing mlp.down_proj in layer 38/50...
+INFO - Start quantizing layer 39/50
+INFO - Quantizing self_attn.k_proj in layer 39/50...
+INFO - Quantizing self_attn.v_proj in layer 39/50...
+INFO - Quantizing self_attn.q_proj in layer 39/50...
+INFO - Quantizing self_attn.o_proj in layer 39/50...
+INFO - Quantizing mlp.up_proj in layer 39/50...
+INFO - Quantizing mlp.gate_proj in layer 39/50...
+INFO - Quantizing mlp.down_proj in layer 39/50...
+INFO - Start quantizing layer 40/50
+INFO - Quantizing self_attn.k_proj in layer 40/50...
+INFO - Quantizing self_attn.v_proj in layer 40/50...
+INFO - Quantizing self_attn.q_proj in layer 40/50...
+INFO - Quantizing self_attn.o_proj in layer 40/50...
+INFO - Quantizing mlp.up_proj in layer 40/50...
+INFO - Quantizing mlp.gate_proj in layer 40/50...
+INFO - Quantizing mlp.down_proj in layer 40/50...
+INFO - Start quantizing layer 41/50
+INFO - Quantizing self_attn.k_proj in layer 41/50...
+INFO - Quantizing self_attn.v_proj in layer 41/50...
+INFO - Quantizing self_attn.q_proj in layer 41/50...
+INFO - Quantizing self_attn.o_proj in layer 41/50...
+INFO - Quantizing mlp.up_proj in layer 41/50...
+INFO - Quantizing mlp.gate_proj in layer 41/50...
+INFO - Quantizing mlp.down_proj in layer 41/50...
+INFO - Start quantizing layer 42/50
+INFO - Quantizing self_attn.k_proj in layer 42/50...
+INFO - Quantizing self_attn.v_proj in layer 42/50...
+INFO - Quantizing self_attn.q_proj in layer 42/50...
+INFO - Quantizing self_attn.o_proj in layer 42/50...
+INFO - Quantizing mlp.up_proj in layer 42/50...
+INFO - Quantizing mlp.gate_proj in layer 42/50...
+INFO - Quantizing mlp.down_proj in layer 42/50...
+INFO - Start quantizing layer 43/50
+INFO - Quantizing self_attn.k_proj in layer 43/50...
+INFO - Quantizing self_attn.v_proj in layer 43/50...
+INFO - Quantizing self_attn.q_proj in layer 43/50...
+INFO - Quantizing self_attn.o_proj in layer 43/50...
+INFO - Quantizing mlp.up_proj in layer 43/50...
+INFO - Quantizing mlp.gate_proj in layer 43/50...
+INFO - Quantizing mlp.down_proj in layer 43/50...
+INFO - Start quantizing layer 44/50
+INFO - Quantizing self_attn.k_proj in layer 44/50...
+INFO - Quantizing self_attn.v_proj in layer 44/50...
+INFO - Quantizing self_attn.q_proj in layer 44/50...
+INFO - Quantizing self_attn.o_proj in layer 44/50...
+INFO - Quantizing mlp.up_proj in layer 44/50...
+INFO - Quantizing mlp.gate_proj in layer 44/50...
+INFO - Quantizing mlp.down_proj in layer 44/50...
+INFO - Start quantizing layer 45/50
+INFO - Quantizing self_attn.k_proj in layer 45/50...
+INFO - Quantizing self_attn.v_proj in layer 45/50...
+INFO - Quantizing self_attn.q_proj in layer 45/50...
+INFO - Quantizing self_attn.o_proj in layer 45/50...
+INFO - Quantizing mlp.up_proj in layer 45/50...
+INFO - Quantizing mlp.gate_proj in layer 45/50...
+INFO - Quantizing mlp.down_proj in layer 45/50...
+INFO - Start quantizing layer 46/50
+INFO - Quantizing self_attn.k_proj in layer 46/50...
+INFO - Quantizing self_attn.v_proj in layer 46/50...
+INFO - Quantizing self_attn.q_proj in layer 46/50...
+INFO - Quantizing self_attn.o_proj in layer 46/50...
+INFO - Quantizing mlp.up_proj in layer 46/50...
+INFO - Quantizing mlp.gate_proj in layer 46/50...
+INFO - Quantizing mlp.down_proj in layer 46/50...
+INFO - Start quantizing layer 47/50
+INFO - Quantizing self_attn.k_proj in layer 47/50...
+INFO - Quantizing self_attn.v_proj in layer 47/50...
+INFO - Quantizing self_attn.q_proj in layer 47/50...
+INFO - Quantizing self_attn.o_proj in layer 47/50...
+INFO - Quantizing mlp.up_proj in layer 47/50...
+INFO - Quantizing mlp.gate_proj in layer 47/50...
+INFO - Quantizing mlp.down_proj in layer 47/50...
+INFO - Start quantizing layer 48/50
+INFO - Quantizing self_attn.k_proj in layer 48/50...
+INFO - Quantizing self_attn.v_proj in layer 48/50...
+INFO - Quantizing self_attn.q_proj in layer 48/50...
+INFO - Quantizing self_attn.o_proj in layer 48/50...
+INFO - Quantizing mlp.up_proj in layer 48/50...
+INFO - Quantizing mlp.gate_proj in layer 48/50...
+INFO - Quantizing mlp.down_proj in layer 48/50...
+INFO - Start quantizing layer 49/50
+INFO - Quantizing self_attn.k_proj in layer 49/50...
+INFO - Quantizing self_attn.v_proj in layer 49/50...
+INFO - Quantizing self_attn.q_proj in layer 49/50...
+INFO - Quantizing self_attn.o_proj in layer 49/50...
+INFO - Quantizing mlp.up_proj in layer 49/50...
+INFO - Quantizing mlp.gate_proj in layer 49/50...
+INFO - Quantizing mlp.down_proj in layer 49/50...
+INFO - Start quantizing layer 50/50
+INFO - Quantizing self_attn.k_proj in layer 50/50...
+INFO - Quantizing self_attn.v_proj in layer 50/50...
+INFO - Quantizing self_attn.q_proj in layer 50/50...
+INFO - Quantizing self_attn.o_proj in layer 50/50...
+INFO - Quantizing mlp.up_proj in layer 50/50...
+INFO - Quantizing mlp.gate_proj in layer 50/50...
+INFO - Quantizing mlp.down_proj in layer 50/50...
+Terminated

variant_b/logs/step4b_v2.log

@@ -0,0 +1,466 @@
+/usr/local/lib/python3.12/dist-packages/auto_gptq/nn_modules/triton_utils/kernels.py:410: FutureWarning: `torch.cuda.amp.custom_fwd(args...)` is deprecated. Please use `torch.amp.custom_fwd(args..., device_type='cuda')` instead.
+  @custom_fwd
+/usr/local/lib/python3.12/dist-packages/auto_gptq/nn_modules/triton_utils/kernels.py:418: FutureWarning: `torch.cuda.amp.custom_bwd(args...)` is deprecated. Please use `torch.amp.custom_bwd(args..., device_type='cuda')` instead.
+  @custom_bwd
+/usr/local/lib/python3.12/dist-packages/auto_gptq/nn_modules/triton_utils/kernels.py:461: FutureWarning: `torch.cuda.amp.custom_fwd(args...)` is deprecated. Please use `torch.amp.custom_fwd(args..., device_type='cuda')` instead.
+  @custom_fwd(cast_inputs=torch.float16)
+CUDA extension not installed.
+CUDA extension not installed.
+Loading calibration data...
+Calibration: 128 samples
+GPTQ config: bits=2, group_size=128, desc_act=False, sym=True
+Loading rotated model for GPTQ...
+
+Moving all model buffers to CUDA...
+  Moving buffer model.layers.0.self_attn.rotary_emb.inv_freq from CPU to CUDA
+  Moving buffer model.layers.1.self_attn.rotary_emb.inv_freq from CPU to CUDA
+  Moving buffer model.layers.2.self_attn.rotary_emb.inv_freq from CPU to CUDA
+  Moving buffer model.layers.3.self_attn.rotary_emb.inv_freq from CPU to CUDA
+  Moving buffer model.layers.4.self_attn.rotary_emb.inv_freq from CPU to CUDA
+  Moving buffer model.layers.5.self_attn.rotary_emb.inv_freq from CPU to CUDA
+  Moving buffer model.layers.6.self_attn.rotary_emb.inv_freq from CPU to CUDA
+  Moving buffer model.layers.7.self_attn.rotary_emb.inv_freq from CPU to CUDA
+  Moving buffer model.layers.8.self_attn.rotary_emb.inv_freq from CPU to CUDA
+  Moving buffer model.layers.9.self_attn.rotary_emb.inv_freq from CPU to CUDA
+  Moving buffer model.layers.10.self_attn.rotary_emb.inv_freq from CPU to CUDA
+  Moving buffer model.layers.11.self_attn.rotary_emb.inv_freq from CPU to CUDA
+  Moving buffer model.layers.12.self_attn.rotary_emb.inv_freq from CPU to CUDA
+  Moving buffer model.layers.13.self_attn.rotary_emb.inv_freq from CPU to CUDA
+  Moving buffer model.layers.14.self_attn.rotary_emb.inv_freq from CPU to CUDA
+  Moving buffer model.layers.15.self_attn.rotary_emb.inv_freq from CPU to CUDA
+  Moving buffer model.layers.16.self_attn.rotary_emb.inv_freq from CPU to CUDA
+  Moving buffer model.layers.17.self_attn.rotary_emb.inv_freq from CPU to CUDA
+  Moving buffer model.layers.18.self_attn.rotary_emb.inv_freq from CPU to CUDA
+  Moving buffer model.layers.19.self_attn.rotary_emb.inv_freq from CPU to CUDA
+  Moving buffer model.layers.20.self_attn.rotary_emb.inv_freq from CPU to CUDA
+  Moving buffer model.layers.21.self_attn.rotary_emb.inv_freq from CPU to CUDA
+  Moving buffer model.layers.22.self_attn.rotary_emb.inv_freq from CPU to CUDA
+  Moving buffer model.layers.23.self_attn.rotary_emb.inv_freq from CPU to CUDA
+  Moving buffer model.layers.24.self_attn.rotary_emb.inv_freq from CPU to CUDA
+  Moving buffer model.layers.25.self_attn.rotary_emb.inv_freq from CPU to CUDA
+  Moving buffer model.layers.26.self_attn.rotary_emb.inv_freq from CPU to CUDA
+  Moving buffer model.layers.27.self_attn.rotary_emb.inv_freq from CPU to CUDA
+  Moving buffer model.layers.28.self_attn.rotary_emb.inv_freq from CPU to CUDA
+  Moving buffer model.layers.29.self_attn.rotary_emb.inv_freq from CPU to CUDA
+  Moving buffer model.layers.30.self_attn.rotary_emb.inv_freq from CPU to CUDA
+  Moving buffer model.layers.31.self_attn.rotary_emb.inv_freq from CPU to CUDA
+  Moving buffer model.layers.32.self_attn.rotary_emb.inv_freq from CPU to CUDA
+  Moving buffer model.layers.33.self_attn.rotary_emb.inv_freq from CPU to CUDA
+  Moving buffer model.layers.34.self_attn.rotary_emb.inv_freq from CPU to CUDA
+  Moving buffer model.layers.35.self_attn.rotary_emb.inv_freq from CPU to CUDA
+  Moving buffer model.layers.36.self_attn.rotary_emb.inv_freq from CPU to CUDA
+  Moving buffer model.layers.37.self_attn.rotary_emb.inv_freq from CPU to CUDA
+  Moving buffer model.layers.38.self_attn.rotary_emb.inv_freq from CPU to CUDA
+  Moving buffer model.layers.39.self_attn.rotary_emb.inv_freq from CPU to CUDA
+  Moving buffer model.layers.40.self_attn.rotary_emb.inv_freq from CPU to CUDA
+  Moving buffer model.layers.41.self_attn.rotary_emb.inv_freq from CPU to CUDA
+  Moving buffer model.layers.42.self_attn.rotary_emb.inv_freq from CPU to CUDA
+  Moving buffer model.layers.43.self_attn.rotary_emb.inv_freq from CPU to CUDA
+  Moving buffer model.layers.44.self_attn.rotary_emb.inv_freq from CPU to CUDA
+  Moving buffer model.layers.45.self_attn.rotary_emb.inv_freq from CPU to CUDA
+  Moving buffer model.layers.46.self_attn.rotary_emb.inv_freq from CPU to CUDA
+  Moving buffer model.layers.47.self_attn.rotary_emb.inv_freq from CPU to CUDA
+  Moving buffer model.layers.48.self_attn.rotary_emb.inv_freq from CPU to CUDA
+  Moving buffer model.layers.49.self_attn.rotary_emb.inv_freq from CPU to CUDA
+  Moving buffer model.rotary_emb.inv_freq from CPU to CUDA
+Running GPTQ 2-bit quantization...
+INFO - Start quantizing layer 1/50
+INFO - Quantizing self_attn.k_proj in layer 1/50...
+INFO - Quantizing self_attn.v_proj in layer 1/50...
+INFO - Quantizing self_attn.q_proj in layer 1/50...
+INFO - Quantizing self_attn.o_proj in layer 1/50...
+INFO - Quantizing mlp.up_proj in layer 1/50...
+INFO - Quantizing mlp.gate_proj in layer 1/50...
+INFO - Quantizing mlp.down_proj in layer 1/50...
+INFO - Start quantizing layer 2/50
+INFO - Quantizing self_attn.k_proj in layer 2/50...
+INFO - Quantizing self_attn.v_proj in layer 2/50...
+INFO - Quantizing self_attn.q_proj in layer 2/50...
+INFO - Quantizing self_attn.o_proj in layer 2/50...
+INFO - Quantizing mlp.up_proj in layer 2/50...
+INFO - Quantizing mlp.gate_proj in layer 2/50...
+INFO - Quantizing mlp.down_proj in layer 2/50...
+INFO - Start quantizing layer 3/50
+INFO - Quantizing self_attn.k_proj in layer 3/50...
+INFO - Quantizing self_attn.v_proj in layer 3/50...
+INFO - Quantizing self_attn.q_proj in layer 3/50...
+INFO - Quantizing self_attn.o_proj in layer 3/50...
+INFO - Quantizing mlp.up_proj in layer 3/50...
+INFO - Quantizing mlp.gate_proj in layer 3/50...
+INFO - Quantizing mlp.down_proj in layer 3/50...
+INFO - Start quantizing layer 4/50
+INFO - Quantizing self_attn.k_proj in layer 4/50...
+INFO - Quantizing self_attn.v_proj in layer 4/50...
+INFO - Quantizing self_attn.q_proj in layer 4/50...
+INFO - Quantizing self_attn.o_proj in layer 4/50...
+INFO - Quantizing mlp.up_proj in layer 4/50...
+INFO - Quantizing mlp.gate_proj in layer 4/50...
+INFO - Quantizing mlp.down_proj in layer 4/50...
+INFO - Start quantizing layer 5/50
+INFO - Quantizing self_attn.k_proj in layer 5/50...
+INFO - Quantizing self_attn.v_proj in layer 5/50...
+INFO - Quantizing self_attn.q_proj in layer 5/50...
+INFO - Quantizing self_attn.o_proj in layer 5/50...
+INFO - Quantizing mlp.up_proj in layer 5/50...
+INFO - Quantizing mlp.gate_proj in layer 5/50...
+INFO - Quantizing mlp.down_proj in layer 5/50...
+INFO - Start quantizing layer 6/50
+INFO - Quantizing self_attn.k_proj in layer 6/50...
+INFO - Quantizing self_attn.v_proj in layer 6/50...
+INFO - Quantizing self_attn.q_proj in layer 6/50...
+INFO - Quantizing self_attn.o_proj in layer 6/50...
+INFO - Quantizing mlp.up_proj in layer 6/50...
+INFO - Quantizing mlp.gate_proj in layer 6/50...
+INFO - Quantizing mlp.down_proj in layer 6/50...
+INFO - Start quantizing layer 7/50
+INFO - Quantizing self_attn.k_proj in layer 7/50...
+INFO - Quantizing self_attn.v_proj in layer 7/50...
+INFO - Quantizing self_attn.q_proj in layer 7/50...
+INFO - Quantizing self_attn.o_proj in layer 7/50...
+INFO - Quantizing mlp.up_proj in layer 7/50...
+INFO - Quantizing mlp.gate_proj in layer 7/50...
+INFO - Quantizing mlp.down_proj in layer 7/50...
+INFO - Start quantizing layer 8/50
+INFO - Quantizing self_attn.k_proj in layer 8/50...
+INFO - Quantizing self_attn.v_proj in layer 8/50...
+INFO - Quantizing self_attn.q_proj in layer 8/50...
+INFO - Quantizing self_attn.o_proj in layer 8/50...
+INFO - Quantizing mlp.up_proj in layer 8/50...
+INFO - Quantizing mlp.gate_proj in layer 8/50...
+INFO - Quantizing mlp.down_proj in layer 8/50...
+INFO - Start quantizing layer 9/50
+INFO - Quantizing self_attn.k_proj in layer 9/50...
+INFO - Quantizing self_attn.v_proj in layer 9/50...
+INFO - Quantizing self_attn.q_proj in layer 9/50...
+INFO - Quantizing self_attn.o_proj in layer 9/50...
+INFO - Quantizing mlp.up_proj in layer 9/50...
+INFO - Quantizing mlp.gate_proj in layer 9/50...
+INFO - Quantizing mlp.down_proj in layer 9/50...
+INFO - Start quantizing layer 10/50
+INFO - Quantizing self_attn.k_proj in layer 10/50...
+INFO - Quantizing self_attn.v_proj in layer 10/50...
+INFO - Quantizing self_attn.q_proj in layer 10/50...
+INFO - Quantizing self_attn.o_proj in layer 10/50...
+INFO - Quantizing mlp.up_proj in layer 10/50...
+INFO - Quantizing mlp.gate_proj in layer 10/50...
+INFO - Quantizing mlp.down_proj in layer 10/50...
+INFO - Start quantizing layer 11/50
+INFO - Quantizing self_attn.k_proj in layer 11/50...
+INFO - Quantizing self_attn.v_proj in layer 11/50...
+INFO - Quantizing self_attn.q_proj in layer 11/50...
+INFO - Quantizing self_attn.o_proj in layer 11/50...
+INFO - Quantizing mlp.up_proj in layer 11/50...
+INFO - Quantizing mlp.gate_proj in layer 11/50...
+INFO - Quantizing mlp.down_proj in layer 11/50...
+INFO - Start quantizing layer 12/50
+INFO - Quantizing self_attn.k_proj in layer 12/50...
+INFO - Quantizing self_attn.v_proj in layer 12/50...
+INFO - Quantizing self_attn.q_proj in layer 12/50...
+INFO - Quantizing self_attn.o_proj in layer 12/50...
+INFO - Quantizing mlp.up_proj in layer 12/50...
+INFO - Quantizing mlp.gate_proj in layer 12/50...
+INFO - Quantizing mlp.down_proj in layer 12/50...
+INFO - Start quantizing layer 13/50
+INFO - Quantizing self_attn.k_proj in layer 13/50...
+INFO - Quantizing self_attn.v_proj in layer 13/50...
+INFO - Quantizing self_attn.q_proj in layer 13/50...
+INFO - Quantizing self_attn.o_proj in layer 13/50...
+INFO - Quantizing mlp.up_proj in layer 13/50...
+INFO - Quantizing mlp.gate_proj in layer 13/50...
+INFO - Quantizing mlp.down_proj in layer 13/50...
+INFO - Start quantizing layer 14/50
+INFO - Quantizing self_attn.k_proj in layer 14/50...
+INFO - Quantizing self_attn.v_proj in layer 14/50...
+INFO - Quantizing self_attn.q_proj in layer 14/50...
+INFO - Quantizing self_attn.o_proj in layer 14/50...
+INFO - Quantizing mlp.up_proj in layer 14/50...
+INFO - Quantizing mlp.gate_proj in layer 14/50...
+INFO - Quantizing mlp.down_proj in layer 14/50...
+INFO - Start quantizing layer 15/50
+INFO - Quantizing self_attn.k_proj in layer 15/50...
+INFO - Quantizing self_attn.v_proj in layer 15/50...
+INFO - Quantizing self_attn.q_proj in layer 15/50...
+INFO - Quantizing self_attn.o_proj in layer 15/50...
+INFO - Quantizing mlp.up_proj in layer 15/50...
+INFO - Quantizing mlp.gate_proj in layer 15/50...
+INFO - Quantizing mlp.down_proj in layer 15/50...
+INFO - Start quantizing layer 16/50
+INFO - Quantizing self_attn.k_proj in layer 16/50...
+INFO - Quantizing self_attn.v_proj in layer 16/50...
+INFO - Quantizing self_attn.q_proj in layer 16/50...
+INFO - Quantizing self_attn.o_proj in layer 16/50...
+INFO - Quantizing mlp.up_proj in layer 16/50...
+INFO - Quantizing mlp.gate_proj in layer 16/50...
+INFO - Quantizing mlp.down_proj in layer 16/50...
+INFO - Start quantizing layer 17/50
+INFO - Quantizing self_attn.k_proj in layer 17/50...
+INFO - Quantizing self_attn.v_proj in layer 17/50...
+INFO - Quantizing self_attn.q_proj in layer 17/50...
+INFO - Quantizing self_attn.o_proj in layer 17/50...
+INFO - Quantizing mlp.up_proj in layer 17/50...
+INFO - Quantizing mlp.gate_proj in layer 17/50...
+INFO - Quantizing mlp.down_proj in layer 17/50...
+INFO - Start quantizing layer 18/50
+INFO - Quantizing self_attn.k_proj in layer 18/50...
+INFO - Quantizing self_attn.v_proj in layer 18/50...
+INFO - Quantizing self_attn.q_proj in layer 18/50...
+INFO - Quantizing self_attn.o_proj in layer 18/50...
+INFO - Quantizing mlp.up_proj in layer 18/50...
+INFO - Quantizing mlp.gate_proj in layer 18/50...
+INFO - Quantizing mlp.down_proj in layer 18/50...
+INFO - Start quantizing layer 19/50
+INFO - Quantizing self_attn.k_proj in layer 19/50...
+INFO - Quantizing self_attn.v_proj in layer 19/50...
+INFO - Quantizing self_attn.q_proj in layer 19/50...
+INFO - Quantizing self_attn.o_proj in layer 19/50...
+INFO - Quantizing mlp.up_proj in layer 19/50...
+INFO - Quantizing mlp.gate_proj in layer 19/50...
+INFO - Quantizing mlp.down_proj in layer 19/50...
+INFO - Start quantizing layer 20/50
+INFO - Quantizing self_attn.k_proj in layer 20/50...
+INFO - Quantizing self_attn.v_proj in layer 20/50...
+INFO - Quantizing self_attn.q_proj in layer 20/50...
+INFO - Quantizing self_attn.o_proj in layer 20/50...
+INFO - Quantizing mlp.up_proj in layer 20/50...
+INFO - Quantizing mlp.gate_proj in layer 20/50...
+INFO - Quantizing mlp.down_proj in layer 20/50...
+INFO - Start quantizing layer 21/50
+INFO - Quantizing self_attn.k_proj in layer 21/50...
+INFO - Quantizing self_attn.v_proj in layer 21/50...
+INFO - Quantizing self_attn.q_proj in layer 21/50...
+INFO - Quantizing self_attn.o_proj in layer 21/50...
+INFO - Quantizing mlp.up_proj in layer 21/50...
+INFO - Quantizing mlp.gate_proj in layer 21/50...
+INFO - Quantizing mlp.down_proj in layer 21/50...
+INFO - Start quantizing layer 22/50
+INFO - Quantizing self_attn.k_proj in layer 22/50...
+INFO - Quantizing self_attn.v_proj in layer 22/50...
+INFO - Quantizing self_attn.q_proj in layer 22/50...
+INFO - Quantizing self_attn.o_proj in layer 22/50...
+INFO - Quantizing mlp.up_proj in layer 22/50...
+INFO - Quantizing mlp.gate_proj in layer 22/50...
+INFO - Quantizing mlp.down_proj in layer 22/50...
+INFO - Start quantizing layer 23/50
+INFO - Quantizing self_attn.k_proj in layer 23/50...
+INFO - Quantizing self_attn.v_proj in layer 23/50...
+INFO - Quantizing self_attn.q_proj in layer 23/50...
+INFO - Quantizing self_attn.o_proj in layer 23/50...
+INFO - Quantizing mlp.up_proj in layer 23/50...
+INFO - Quantizing mlp.gate_proj in layer 23/50...
+INFO - Quantizing mlp.down_proj in layer 23/50...
+INFO - Start quantizing layer 24/50
+INFO - Quantizing self_attn.k_proj in layer 24/50...
+INFO - Quantizing self_attn.v_proj in layer 24/50...
+INFO - Quantizing self_attn.q_proj in layer 24/50...
+INFO - Quantizing self_attn.o_proj in layer 24/50...
+INFO - Quantizing mlp.up_proj in layer 24/50...
+INFO - Quantizing mlp.gate_proj in layer 24/50...
+INFO - Quantizing mlp.down_proj in layer 24/50...
+INFO - Start quantizing layer 25/50
+INFO - Quantizing self_attn.k_proj in layer 25/50...
+INFO - Quantizing self_attn.v_proj in layer 25/50...
+INFO - Quantizing self_attn.q_proj in layer 25/50...
+INFO - Quantizing self_attn.o_proj in layer 25/50...
+INFO - Quantizing mlp.up_proj in layer 25/50...
+INFO - Quantizing mlp.gate_proj in layer 25/50...
+INFO - Quantizing mlp.down_proj in layer 25/50...
+INFO - Start quantizing layer 26/50
+INFO - Quantizing self_attn.k_proj in layer 26/50...
+INFO - Quantizing self_attn.v_proj in layer 26/50...
+INFO - Quantizing self_attn.q_proj in layer 26/50...
+INFO - Quantizing self_attn.o_proj in layer 26/50...
+INFO - Quantizing mlp.up_proj in layer 26/50...
+INFO - Quantizing mlp.gate_proj in layer 26/50...
+INFO - Quantizing mlp.down_proj in layer 26/50...
+INFO - Start quantizing layer 27/50
+INFO - Quantizing self_attn.k_proj in layer 27/50...
+INFO - Quantizing self_attn.v_proj in layer 27/50...
+INFO - Quantizing self_attn.q_proj in layer 27/50...
+INFO - Quantizing self_attn.o_proj in layer 27/50...
+INFO - Quantizing mlp.up_proj in layer 27/50...
+INFO - Quantizing mlp.gate_proj in layer 27/50...
+INFO - Quantizing mlp.down_proj in layer 27/50...
+INFO - Start quantizing layer 28/50
+INFO - Quantizing self_attn.k_proj in layer 28/50...
+INFO - Quantizing self_attn.v_proj in layer 28/50...
+INFO - Quantizing self_attn.q_proj in layer 28/50...
+INFO - Quantizing self_attn.o_proj in layer 28/50...
+INFO - Quantizing mlp.up_proj in layer 28/50...
+INFO - Quantizing mlp.gate_proj in layer 28/50...
+INFO - Quantizing mlp.down_proj in layer 28/50...
+INFO - Start quantizing layer 29/50
+INFO - Quantizing self_attn.k_proj in layer 29/50...
+INFO - Quantizing self_attn.v_proj in layer 29/50...
+INFO - Quantizing self_attn.q_proj in layer 29/50...
+INFO - Quantizing self_attn.o_proj in layer 29/50...
+INFO - Quantizing mlp.up_proj in layer 29/50...
+INFO - Quantizing mlp.gate_proj in layer 29/50...
+INFO - Quantizing mlp.down_proj in layer 29/50...
+INFO - Start quantizing layer 30/50
+INFO - Quantizing self_attn.k_proj in layer 30/50...
+INFO - Quantizing self_attn.v_proj in layer 30/50...
+INFO - Quantizing self_attn.q_proj in layer 30/50...
+INFO - Quantizing self_attn.o_proj in layer 30/50...
+INFO - Quantizing mlp.up_proj in layer 30/50...
+INFO - Quantizing mlp.gate_proj in layer 30/50...
+INFO - Quantizing mlp.down_proj in layer 30/50...
+INFO - Start quantizing layer 31/50
+INFO - Quantizing self_attn.k_proj in layer 31/50...
+INFO - Quantizing self_attn.v_proj in layer 31/50...
+INFO - Quantizing self_attn.q_proj in layer 31/50...
+INFO - Quantizing self_attn.o_proj in layer 31/50...
+INFO - Quantizing mlp.up_proj in layer 31/50...
+INFO - Quantizing mlp.gate_proj in layer 31/50...
+INFO - Quantizing mlp.down_proj in layer 31/50...
+INFO - Start quantizing layer 32/50
+INFO - Quantizing self_attn.k_proj in layer 32/50...
+INFO - Quantizing self_attn.v_proj in layer 32/50...
+INFO - Quantizing self_attn.q_proj in layer 32/50...
+INFO - Quantizing self_attn.o_proj in layer 32/50...
+INFO - Quantizing mlp.up_proj in layer 32/50...
+INFO - Quantizing mlp.gate_proj in layer 32/50...
+INFO - Quantizing mlp.down_proj in layer 32/50...
+INFO - Start quantizing layer 33/50
+INFO - Quantizing self_attn.k_proj in layer 33/50...
+INFO - Quantizing self_attn.v_proj in layer 33/50...
+INFO - Quantizing self_attn.q_proj in layer 33/50...
+INFO - Quantizing self_attn.o_proj in layer 33/50...
+INFO - Quantizing mlp.up_proj in layer 33/50...
+INFO - Quantizing mlp.gate_proj in layer 33/50...
+INFO - Quantizing mlp.down_proj in layer 33/50...
+INFO - Start quantizing layer 34/50
+INFO - Quantizing self_attn.k_proj in layer 34/50...
+INFO - Quantizing self_attn.v_proj in layer 34/50...
+INFO - Quantizing self_attn.q_proj in layer 34/50...
+INFO - Quantizing self_attn.o_proj in layer 34/50...
+INFO - Quantizing mlp.up_proj in layer 34/50...
+INFO - Quantizing mlp.gate_proj in layer 34/50...
+INFO - Quantizing mlp.down_proj in layer 34/50...
+INFO - Start quantizing layer 35/50
+INFO - Quantizing self_attn.k_proj in layer 35/50...
+INFO - Quantizing self_attn.v_proj in layer 35/50...
+INFO - Quantizing self_attn.q_proj in layer 35/50...
+INFO - Quantizing self_attn.o_proj in layer 35/50...
+INFO - Quantizing mlp.up_proj in layer 35/50...
+INFO - Quantizing mlp.gate_proj in layer 35/50...
+INFO - Quantizing mlp.down_proj in layer 35/50...
+INFO - Start quantizing layer 36/50
+INFO - Quantizing self_attn.k_proj in layer 36/50...
+INFO - Quantizing self_attn.v_proj in layer 36/50...
+INFO - Quantizing self_attn.q_proj in layer 36/50...
+INFO - Quantizing self_attn.o_proj in layer 36/50...
+INFO - Quantizing mlp.up_proj in layer 36/50...
+INFO - Quantizing mlp.gate_proj in layer 36/50...
+INFO - Quantizing mlp.down_proj in layer 36/50...
+INFO - Start quantizing layer 37/50
+INFO - Quantizing self_attn.k_proj in layer 37/50...
+INFO - Quantizing self_attn.v_proj in layer 37/50...
+INFO - Quantizing self_attn.q_proj in layer 37/50...
+INFO - Quantizing self_attn.o_proj in layer 37/50...
+INFO - Quantizing mlp.up_proj in layer 37/50...
+INFO - Quantizing mlp.gate_proj in layer 37/50...
+INFO - Quantizing mlp.down_proj in layer 37/50...
+INFO - Start quantizing layer 38/50
+INFO - Quantizing self_attn.k_proj in layer 38/50...
+INFO - Quantizing self_attn.v_proj in layer 38/50...
+INFO - Quantizing self_attn.q_proj in layer 38/50...
+INFO - Quantizing self_attn.o_proj in layer 38/50...
+INFO - Quantizing mlp.up_proj in layer 38/50...
+INFO - Quantizing mlp.gate_proj in layer 38/50...
+INFO - Quantizing mlp.down_proj in layer 38/50...
+INFO - Start quantizing layer 39/50
+INFO - Quantizing self_attn.k_proj in layer 39/50...
+INFO - Quantizing self_attn.v_proj in layer 39/50...
+INFO - Quantizing self_attn.q_proj in layer 39/50...
+INFO - Quantizing self_attn.o_proj in layer 39/50...
+INFO - Quantizing mlp.up_proj in layer 39/50...
+INFO - Quantizing mlp.gate_proj in layer 39/50...
+INFO - Quantizing mlp.down_proj in layer 39/50...
+INFO - Start quantizing layer 40/50
+INFO - Quantizing self_attn.k_proj in layer 40/50...
+INFO - Quantizing self_attn.v_proj in layer 40/50...
+INFO - Quantizing self_attn.q_proj in layer 40/50...
+INFO - Quantizing self_attn.o_proj in layer 40/50...
+INFO - Quantizing mlp.up_proj in layer 40/50...
+INFO - Quantizing mlp.gate_proj in layer 40/50...
+INFO - Quantizing mlp.down_proj in layer 40/50...
+INFO - Start quantizing layer 41/50
+INFO - Quantizing self_attn.k_proj in layer 41/50...
+INFO - Quantizing self_attn.v_proj in layer 41/50...
+INFO - Quantizing self_attn.q_proj in layer 41/50...
+INFO - Quantizing self_attn.o_proj in layer 41/50...
+INFO - Quantizing mlp.up_proj in layer 41/50...
+INFO - Quantizing mlp.gate_proj in layer 41/50...
+INFO - Quantizing mlp.down_proj in layer 41/50...
+INFO - Start quantizing layer 42/50
+INFO - Quantizing self_attn.k_proj in layer 42/50...
+INFO - Quantizing self_attn.v_proj in layer 42/50...
+INFO - Quantizing self_attn.q_proj in layer 42/50...
+INFO - Quantizing self_attn.o_proj in layer 42/50...
+INFO - Quantizing mlp.up_proj in layer 42/50...
+INFO - Quantizing mlp.gate_proj in layer 42/50...
+INFO - Quantizing mlp.down_proj in layer 42/50...
+INFO - Start quantizing layer 43/50
+INFO - Quantizing self_attn.k_proj in layer 43/50...
+INFO - Quantizing self_attn.v_proj in layer 43/50...
+INFO - Quantizing self_attn.q_proj in layer 43/50...
+INFO - Quantizing self_attn.o_proj in layer 43/50...
+INFO - Quantizing mlp.up_proj in layer 43/50...
+INFO - Quantizing mlp.gate_proj in layer 43/50...
+INFO - Quantizing mlp.down_proj in layer 43/50...
+INFO - Start quantizing layer 44/50
+INFO - Quantizing self_attn.k_proj in layer 44/50...
+INFO - Quantizing self_attn.v_proj in layer 44/50...
+INFO - Quantizing self_attn.q_proj in layer 44/50...
+INFO - Quantizing self_attn.o_proj in layer 44/50...
+INFO - Quantizing mlp.up_proj in layer 44/50...
+INFO - Quantizing mlp.gate_proj in layer 44/50...
+INFO - Quantizing mlp.down_proj in layer 44/50...
+INFO - Start quantizing layer 45/50
+INFO - Quantizing self_attn.k_proj in layer 45/50...
+INFO - Quantizing self_attn.v_proj in layer 45/50...
+INFO - Quantizing self_attn.q_proj in layer 45/50...
+INFO - Quantizing self_attn.o_proj in layer 45/50...
+INFO - Quantizing mlp.up_proj in layer 45/50...
+INFO - Quantizing mlp.gate_proj in layer 45/50...
+INFO - Quantizing mlp.down_proj in layer 45/50...
+INFO - Start quantizing layer 46/50
+INFO - Quantizing self_attn.k_proj in layer 46/50...
+INFO - Quantizing self_attn.v_proj in layer 46/50...
+INFO - Quantizing self_attn.q_proj in layer 46/50...
+INFO - Quantizing self_attn.o_proj in layer 46/50...
+INFO - Quantizing mlp.up_proj in layer 46/50...
+INFO - Quantizing mlp.gate_proj in layer 46/50...
+INFO - Quantizing mlp.down_proj in layer 46/50...
+INFO - Start quantizing layer 47/50
+INFO - Quantizing self_attn.k_proj in layer 47/50...
+INFO - Quantizing self_attn.v_proj in layer 47/50...
+INFO - Quantizing self_attn.q_proj in layer 47/50...
+INFO - Quantizing self_attn.o_proj in layer 47/50...
+INFO - Quantizing mlp.up_proj in layer 47/50...
+INFO - Quantizing mlp.gate_proj in layer 47/50...
+INFO - Quantizing mlp.down_proj in layer 47/50...
+INFO - Start quantizing layer 48/50
+INFO - Quantizing self_attn.k_proj in layer 48/50...
+INFO - Quantizing self_attn.v_proj in layer 48/50...
+INFO - Quantizing self_attn.q_proj in layer 48/50...
+INFO - Quantizing self_attn.o_proj in layer 48/50...
+INFO - Quantizing mlp.up_proj in layer 48/50...
+INFO - Quantizing mlp.gate_proj in layer 48/50...
+INFO - Quantizing mlp.down_proj in layer 48/50...
+INFO - Start quantizing layer 49/50
+INFO - Quantizing self_attn.k_proj in layer 49/50...
+INFO - Quantizing self_attn.v_proj in layer 49/50...
+INFO - Quantizing self_attn.q_proj in layer 49/50...
+INFO - Quantizing self_attn.o_proj in layer 49/50...
+INFO - Quantizing mlp.up_proj in layer 49/50...
+INFO - Quantizing mlp.gate_proj in layer 49/50...
+INFO - Quantizing mlp.down_proj in layer 49/50...
+INFO - Start quantizing layer 50/50
+INFO - Quantizing self_attn.k_proj in layer 50/50...
+INFO - Quantizing self_attn.v_proj in layer 50/50...
+INFO - Quantizing self_attn.q_proj in layer 50/50...
+INFO - Quantizing self_attn.o_proj in layer 50/50...
+INFO - Quantizing mlp.up_proj in layer 50/50...
+INFO - Quantizing mlp.gate_proj in layer 50/50...
+INFO - Quantizing mlp.down_proj in layer 50/50...

variant_b/logs/step5_mc.log

@@ -0,0 +1,62 @@
+2026-02-22:00:37:14,293 INFO     [__main__.py:272] Verbosity set to INFO
+2026-02-22:00:37:17,278 INFO     [__main__.py:363] Selected Tasks: ['polish_mc']
+2026-02-22:00:37:17,281 INFO     [evaluator.py:152] Setting random seed to 0 | Setting numpy seed to 1234 | Setting torch manual seed to 1234
+2026-02-22:00:37:17,281 INFO     [evaluator.py:189] Initializing hf model, with arguments: {'pretrained': '/dev/shm/spinquant/exported_model', 'dtype': 'bfloat16', 'trust_remote_code': True}
+2026-02-22:00:37:17,643 INFO     [huggingface.py:169] Using device 'cuda'
+/usr/local/lib/python3.12/dist-packages/auto_gptq/nn_modules/triton_utils/kernels.py:410: FutureWarning: `torch.cuda.amp.custom_fwd(args...)` is deprecated. Please use `torch.amp.custom_fwd(args..., device_type='cuda')` instead.
+  @custom_fwd
+/usr/local/lib/python3.12/dist-packages/auto_gptq/nn_modules/triton_utils/kernels.py:418: FutureWarning: `torch.cuda.amp.custom_bwd(args...)` is deprecated. Please use `torch.amp.custom_bwd(args..., device_type='cuda')` instead.
+  @custom_bwd
+/usr/local/lib/python3.12/dist-packages/auto_gptq/nn_modules/triton_utils/kernels.py:461: FutureWarning: `torch.cuda.amp.custom_fwd(args...)` is deprecated. Please use `torch.amp.custom_fwd(args..., device_type='cuda')` instead.
+  @custom_fwd(cast_inputs=torch.float16)
+2026-02-22:00:37:17,897 WARNING  [qlinear_cuda.py:18] CUDA extension not installed.
+2026-02-22:00:37:17,897 WARNING  [qlinear_cuda_old.py:17] CUDA extension not installed.
+/usr/local/lib/python3.12/dist-packages/transformers/modeling_utils.py:4674: FutureWarning: `_is_quantized_training_enabled` is going to be deprecated in transformers 4.39.0. Please use `model.hf_quantizer.is_trainable` instead
+  warnings.warn(
+Traceback (most recent call last):
+  File "/usr/local/bin/lm_eval", line 8, in <module>
+    sys.exit(cli_evaluate())
+             ^^^^^^^^^^^^^^
+  File "/dev/shm/eval/lm-evaluation-harness/lm_eval/__main__.py", line 369, in cli_evaluate
+    results = evaluator.simple_evaluate(
+              ^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/dev/shm/eval/lm-evaluation-harness/lm_eval/utils.py", line 346, in _wrapper
+    return fn(*args, **kwargs)
+           ^^^^^^^^^^^^^^^^^^^
+  File "/dev/shm/eval/lm-evaluation-harness/lm_eval/evaluator.py", line 221, in simple_evaluate
+    task_dict = get_task_dict(tasks, task_manager)
+                ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/dev/shm/eval/lm-evaluation-harness/lm_eval/tasks/__init__.py", line 423, in get_task_dict
+    task_name_from_string_dict = task_manager.load_task_or_group(
+                                 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/dev/shm/eval/lm-evaluation-harness/lm_eval/tasks/__init__.py", line 271, in load_task_or_group
+    collections.ChainMap(*map(self._load_individual_task_or_group, task_list))
+  File "/dev/shm/eval/lm-evaluation-harness/lm_eval/tasks/__init__.py", line 254, in _load_individual_task_or_group
+    **dict(collections.ChainMap(*map(fn, subtask_list))),
+           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/dev/shm/eval/lm-evaluation-harness/lm_eval/tasks/__init__.py", line 162, in _load_individual_task_or_group
+    return load_task(task_config, task=name_or_config, group=parent_name)
+           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/dev/shm/eval/lm-evaluation-harness/lm_eval/tasks/__init__.py", line 151, in load_task
+    task_object = ConfigurableTask(config=config)
+                  ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/dev/shm/eval/lm-evaluation-harness/lm_eval/api/task.py", line 809, in __init__
+    self.download(self.config.dataset_kwargs)
+  File "/dev/shm/eval/lm-evaluation-harness/lm_eval/api/task.py", line 916, in download
+    self.dataset = datasets.load_dataset(
+                   ^^^^^^^^^^^^^^^^^^^^^^
+  File "/usr/local/lib/python3.12/dist-packages/datasets/load.py", line 1488, in load_dataset
+    builder_instance = load_dataset_builder(
+                       ^^^^^^^^^^^^^^^^^^^^^
+  File "/usr/local/lib/python3.12/dist-packages/datasets/load.py", line 1167, in load_dataset_builder
+    builder_instance: DatasetBuilder = builder_cls(
+                                       ^^^^^^^^^^^^
+  File "/usr/local/lib/python3.12/dist-packages/datasets/builder.py", line 343, in __init__
+    self.config, self.config_id = self._create_builder_config(
+                                  ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/usr/local/lib/python3.12/dist-packages/datasets/builder.py", line 515, in _create_builder_config
+    raise ValueError(
+ValueError: Config name is missing.
+Please pick one among the available configs: ['acm_Arab', 'arz_Arab', 'ceb_Latn', 'fin_Latn', 'hin_Deva', 'ita_Latn', 'khm_Khmr', 'lvs_Latn', 'npi_Deva', 'pol_Latn', 'slv_Latn', 'swe_Latn', 'tso_Latn', 'xho_Latn', 'afr_Latn', 'asm_Beng', 'ces_Latn', 'fra_Latn', 'hin_Latn', 'jav_Latn', 'kin_Latn', 'mal_Mlym', 'npi_Latn', 'por_Latn', 'sna_Latn', 'swh_Latn', 'tur_Latn', 'yor_Latn', 'als_Latn', 'azj_Latn', 'ckb_Arab', 'fuv_Latn', 'hrv_Latn', 'jpn_Jpan', 'kir_Cyrl', 'mar_Deva', 'nso_Latn', 'snd_Arab', 'tam_Taml', 'ukr_Cyrl', 'zho_Hans', 'amh_Ethi', 'bam_Latn', 'dan_Latn', 'gaz_Latn', 'hun_Latn', 'kac_Latn', 'kor_Hang', 'mkd_Cyrl', 'nya_Latn', 'ron_Latn', 'som_Latn', 'tel_Telu', 'urd_Arab', 'zho_Hant', 'apc_Arab', 'ben_Beng', 'deu_Latn', 'grn_Latn', 'hye_Armn', 'kan_Knda', 'lao_Laoo', 'mlt_Latn', 'ory_Orya', 'rus_Cyrl', 'sot_Latn', 'tgk_Cyrl', 'urd_Latn', 'zsm_Latn', 'arb_Arab', 'ben_Latn', 'ell_Grek', 'guj_Gujr', 'ibo_Latn', 'kat_Geor', 'lin_Latn', 'mri_Latn', 'pan_Guru', 'shn_Mymr', 'spa_Latn', 'tgl_Latn', 'uzn_Latn', 'zul_Latn', 'arb_Latn', 'bod_Tibt', 'eng_Latn', 'hat_Latn', 'ilo_Latn', 'kaz_Cyrl', 'lit_Latn', 'mya_Mymr', 'pbt_Arab', 'sin_Latn', 'srp_Cyrl', 'tha_Thai', 'vie_Latn', 'ars_Arab', 'bul_Cyrl', 'est_Latn', 'hau_Latn', 'ind_Latn', 'kea_Latn', 'lug_Latn', 'nld_Latn', 'pes_Arab', 'sin_Sinh', 'ssw_Latn', 'tir_Ethi', 'war_Latn', 'ary_Arab', 'cat_Latn', 'eus_Latn', 'heb_Hebr', 'isl_Latn', 'khk_Cyrl', 'luo_Latn', 'nob_Latn', 'plt_Latn', 'slk_Latn', 'sun_Latn', 'tsn_Latn', 'wol_Latn']
+Example of usage:
+	`load_dataset('facebook/belebele', 'acm_Arab')`

variant_b/logs/step5_output.log

@@ -0,0 +1,80 @@
+[2026-02-22 00:30:13] ========== STEP 5: EVALUATION ==========
+[2026-02-22 00:30:13] Setting up lm-evaluation-harness (polish3 branch)...
+    Found existing installation: lm_eval 0.4.2
+    Uninstalling lm_eval-0.4.2:
+      Successfully uninstalled lm_eval-0.4.2
+Successfully installed lm_eval-0.4.2
+WARNING: Running pip as the 'root' user can result in broken permissions and conflicting behaviour with the system package manager. It is recommended to use a virtual environment instead: https://pip.pypa.io/warnings/venv
+[2026-02-22 00:30:20] lm-evaluation-harness installed
+[2026-02-22 00:30:20] Running eval batch 1: polish_mc...
+2026-02-22:00:30:24,979 INFO     [__main__.py:272] Verbosity set to INFO
+2026-02-22:00:30:27,948 INFO     [__main__.py:363] Selected Tasks: ['polish_mc']
+2026-02-22:00:30:27,951 INFO     [evaluator.py:152] Setting random seed to 0 | Setting numpy seed to 1234 | Setting torch manual seed to 1234
+2026-02-22:00:30:27,951 INFO     [evaluator.py:189] Initializing hf model, with arguments: {'pretrained': '/dev/shm/spinquant/exported_model', 'dtype': 'bfloat16', 'trust_remote_code': True}
+2026-02-22:00:30:28,312 INFO     [huggingface.py:169] Using device 'cuda'
+/usr/local/lib/python3.12/dist-packages/auto_gptq/nn_modules/triton_utils/kernels.py:410: FutureWarning: `torch.cuda.amp.custom_fwd(args...)` is deprecated. Please use `torch.amp.custom_fwd(args..., device_type='cuda')` instead.
+  @custom_fwd
+/usr/local/lib/python3.12/dist-packages/auto_gptq/nn_modules/triton_utils/kernels.py:418: FutureWarning: `torch.cuda.amp.custom_bwd(args...)` is deprecated. Please use `torch.amp.custom_bwd(args..., device_type='cuda')` instead.
+  @custom_bwd
+/usr/local/lib/python3.12/dist-packages/auto_gptq/nn_modules/triton_utils/kernels.py:461: FutureWarning: `torch.cuda.amp.custom_fwd(args...)` is deprecated. Please use `torch.amp.custom_fwd(args..., device_type='cuda')` instead.
+  @custom_fwd(cast_inputs=torch.float16)
+2026-02-22:00:30:28,565 WARNING  [qlinear_cuda.py:18] CUDA extension not installed.
+2026-02-22:00:30:28,566 WARNING  [qlinear_cuda_old.py:17] CUDA extension not installed.
+/usr/local/lib/python3.12/dist-packages/transformers/modeling_utils.py:4674: FutureWarning: `_is_quantized_training_enabled` is going to be deprecated in transformers 4.39.0. Please use `model.hf_quantizer.is_trainable` instead
+  warnings.warn(
+
+
+
+
+
+
+
+
+
+Traceback (most recent call last):
+  File "/usr/local/bin/lm_eval", line 8, in <module>
+    sys.exit(cli_evaluate())
+             ^^^^^^^^^^^^^^
+  File "/dev/shm/eval/lm-evaluation-harness/lm_eval/__main__.py", line 369, in cli_evaluate
+    results = evaluator.simple_evaluate(
+              ^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/dev/shm/eval/lm-evaluation-harness/lm_eval/utils.py", line 346, in _wrapper
+    return fn(*args, **kwargs)
+           ^^^^^^^^^^^^^^^^^^^
+  File "/dev/shm/eval/lm-evaluation-harness/lm_eval/evaluator.py", line 221, in simple_evaluate
+    task_dict = get_task_dict(tasks, task_manager)
+                ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/dev/shm/eval/lm-evaluation-harness/lm_eval/tasks/__init__.py", line 423, in get_task_dict
+    task_name_from_string_dict = task_manager.load_task_or_group(
+                                 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/dev/shm/eval/lm-evaluation-harness/lm_eval/tasks/__init__.py", line 271, in load_task_or_group
+    collections.ChainMap(*map(self._load_individual_task_or_group, task_list))
+  File "/dev/shm/eval/lm-evaluation-harness/lm_eval/tasks/__init__.py", line 254, in _load_individual_task_or_group
+    **dict(collections.ChainMap(*map(fn, subtask_list))),
+           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/dev/shm/eval/lm-evaluation-harness/lm_eval/tasks/__init__.py", line 162, in _load_individual_task_or_group
+    return load_task(task_config, task=name_or_config, group=parent_name)
+           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/dev/shm/eval/lm-evaluation-harness/lm_eval/tasks/__init__.py", line 151, in load_task
+    task_object = ConfigurableTask(config=config)
+                  ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/dev/shm/eval/lm-evaluation-harness/lm_eval/api/task.py", line 809, in __init__
+    self.download(self.config.dataset_kwargs)
+  File "/dev/shm/eval/lm-evaluation-harness/lm_eval/api/task.py", line 916, in download
+    self.dataset = datasets.load_dataset(
+                   ^^^^^^^^^^^^^^^^^^^^^^
+  File "/usr/local/lib/python3.12/dist-packages/datasets/load.py", line 1488, in load_dataset
+    builder_instance = load_dataset_builder(
+                       ^^^^^^^^^^^^^^^^^^^^^
+  File "/usr/local/lib/python3.12/dist-packages/datasets/load.py", line 1167, in load_dataset_builder
+    builder_instance: DatasetBuilder = builder_cls(
+                                       ^^^^^^^^^^^^
+  File "/usr/local/lib/python3.12/dist-packages/datasets/builder.py", line 343, in __init__
+    self.config, self.config_id = self._create_builder_config(
+                                  ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+  File "/usr/local/lib/python3.12/dist-packages/datasets/builder.py", line 515, in _create_builder_config
+    raise ValueError(
+ValueError: Config name is missing.
+Please pick one among the available configs: ['acm_Arab', 'arz_Arab', 'ceb_Latn', 'fin_Latn', 'hin_Deva', 'ita_Latn', 'khm_Khmr', 'lvs_Latn', 'npi_Deva', 'pol_Latn', 'slv_Latn', 'swe_Latn', 'tso_Latn', 'xho_Latn', 'afr_Latn', 'asm_Beng', 'ces_Latn', 'fra_Latn', 'hin_Latn', 'jav_Latn', 'kin_Latn', 'mal_Mlym', 'npi_Latn', 'por_Latn', 'sna_Latn', 'swh_Latn', 'tur_Latn', 'yor_Latn', 'als_Latn', 'azj_Latn', 'ckb_Arab', 'fuv_Latn', 'hrv_Latn', 'jpn_Jpan', 'kir_Cyrl', 'mar_Deva', 'nso_Latn', 'snd_Arab', 'tam_Taml', 'ukr_Cyrl', 'zho_Hans', 'amh_Ethi', 'bam_Latn', 'dan_Latn', 'gaz_Latn', 'hun_Latn', 'kac_Latn', 'kor_Hang', 'mkd_Cyrl', 'nya_Latn', 'ron_Latn', 'som_Latn', 'tel_Telu', 'urd_Arab', 'zho_Hant', 'apc_Arab', 'ben_Beng', 'deu_Latn', 'grn_Latn', 'hye_Armn', 'kan_Knda', 'lao_Laoo', 'mlt_Latn', 'ory_Orya', 'rus_Cyrl', 'sot_Latn', 'tgk_Cyrl', 'urd_Latn', 'zsm_Latn', 'arb_Arab', 'ben_Latn', 'ell_Grek', 'guj_Gujr', 'ibo_Latn', 'kat_Geor', 'lin_Latn', 'mri_Latn', 'pan_Guru', 'shn_Mymr', 'spa_Latn', 'tgl_Latn', 'uzn_Latn', 'zul_Latn', 'arb_Latn', 'bod_Tibt', 'eng_Latn', 'hat_Latn', 'ilo_Latn', 'kaz_Cyrl', 'lit_Latn', 'mya_Mymr', 'pbt_Arab', 'sin_Latn', 'srp_Cyrl', 'tha_Thai', 'vie_Latn', 'ars_Arab', 'bul_Cyrl', 'est_Latn', 'hau_Latn', 'ind_Latn', 'kea_Latn', 'lug_Latn', 'nld_Latn', 'pes_Arab', 'sin_Sinh', 'ssw_Latn', 'tir_Ethi', 'war_Latn', 'ary_Arab', 'cat_Latn', 'eus_Latn', 'heb_Hebr', 'isl_Latn', 'khk_Cyrl', 'luo_Latn', 'nob_Latn', 'plt_Latn', 'slk_Latn', 'sun_Latn', 'tsn_Latn', 'wol_Latn']
+Example of usage:
+	`load_dataset('facebook/belebele', 'acm_Arab')`

variant_b/rbin_info.txt

@@ -0,0 +1,2 @@
+R.bin size: 70398323 bytes (67.1 MB)
+Path: /workspace/R.bin (backup) + rotations/R.bin in model repo

variant_b/report/variant_b_summary.md

@@ -0,0 +1,126 @@
+# Variant B: SpinQuant + GPTQ 2-bit — Podsumowanie
+
+## 1. Co zrobiliśmy
+
+### Pipeline
+1. **Model bazowy:** `speakleash/Bielik-11B-v2.3-Instruct` (Mistral 7B arch, 50 warstw, 4096 hidden, 32 heads, 8 KV heads)
+2. **SpinQuant rotacja** (offline, na H200 GPU):
+   - `fuse_layer_norms()` — wycentrowanie embeddingów (odjęcie średniej), fuzja wag LayerNorm do sąsiednich warstw liniowych, ustawienie wszystkich LN weights = 1.0
+   - **R1** (4096×4096 macierz ortogonalna): `embed @ R1`, `q/k/v_proj @ R1`, `R1.T @ o_proj`, `up/gate_proj @ R1`, `R1.T @ down_proj`, `lm_head @ R1`
+   - **R2** (50× 128×128 macierzy ortogonalnych, per-layer): na `v_proj` (output side) i `o_proj` (input side) per-head
+   - **Pominięto:** transformatę Hadamarda na `down_proj` (wymaga runtime Hadamard na aktywacjach, nieobsługiwany przez gptqmodel)
+3. **Weryfikacja rotowanego modelu FP16** — generacja poprawnego tekstu po polsku
+4. **GPTQ kwantyzacja 2-bit** (gptqmodel 5.7.0 na H200):
+   - Konfiguracja: `bits=2, group_size=128, sym=True, damp_percent=0.01`
+   - Kalibracja: Polish Wikipedia (`wikimedia/wikipedia 20231101.pl`), 128–256 próbek × 2048 tokenów
+   - Testowano: `desc_act=True` i `desc_act=False`
+5. **Weryfikacja kwantyzowanego modelu** — analiza logitów i open generation
+
+### Narzędzia
+- **GPU:** NVIDIA H200 80GB (vast.ai)
+- **Software:** PyTorch 2.10+cu128, transformers 5.2.0, gptqmodel 5.7.0
+- **Rotacje:** R1 i R2 z pliku `R.bin` (wygenerowane wcześniej przez SpinQuant)
+- **Kod referencyjny:** https://github.com/facebookresearch/SpinQuant
+
+### Repozytoria HuggingFace
+| Repo | Opis | Status |
+|------|-------|--------|
+| `Jakubrd4/bielik-q2-variant-b` | Oryginał (SpinQuant+GPTQ, desc_act=False) | Nietknięty |
+| `Jakubrd4/bielik-q2-variant-b-fixed` | Poprawiony config.json (mistral zamiast llama) | Prywatny |
+
+---
+
+## 2. Wyniki
+
+### Config fix (model_type llama → mistral)
+Oryginalny model miał `model_type: "llama"` i `architectures: ["LlamaForCausalLM"]` zamiast poprawnych `"mistral"` / `["MistralForCausalLM"]`. Naprawienie config **nie rozwiązało** problemu z generacją — oba warianty generowały identyczne bzdury.
+
+### Rotowany model FP16 (przed kwantyzacją)
+| Prompt | Top-1 token | Prawdopodobieństwo | Entropia |
+|--------|-------------|-------------------|----------|
+| "Warszawa jest stolica" | "Pol" | 70.96% | 1.61 |
+| "Najwyzsza gora w Polsce to" | "R" (Rysy) | 89% | ~0.8 |
+
+**Generacja:** Poprawny, płynny tekst po polsku. Model FP16 po rotacji działa idealnie.
+
+### Kwantyzowany model 2-bit GPTQ (desc_act=True)
+| Prompt | Top-1 token | Prawdopodobieństwo | Entropia |
+|--------|-------------|-------------------|----------|
+| "Warszawa jest stolica" | "ex" | 7.28% | 6.12 |
+
+**Generacja:** `"Warszawa jest stolica ex to wiwe se toi tozabwę autorskaw ęsywęc do"` — bzdury z polskimi znakami.
+
+### Kwantyzowany model 2-bit GPTQ (desc_act=False) — oryginalny z HF
+| Metryka | Wartość |
+|---------|---------|
+| DYK accuracy (MC eval) | 62.88% |
+| Open generation | Bzdury (`"erneRegistryRegistry..."`) |
+| Logity: entropia | 9.00 |
+| Logity: max prawdopodobieństwo | 0.78% |
+
+### Porównanie entropii (niżej = lepiej)
+```
+FP16 rotowany:     1.61  ████
+2-bit desc_act=T:  6.12  ████████████████████
+2-bit desc_act=F:  9.00  █████████████████████████████ (oryginalny z HF)
+Losowe:           10.37  █████████████████████████████████
+```
+
+---
+
+## 3. Problemy napotkane
+
+### Problem 1: Błędny config.json
+- **Objaw:** `model_type: "llama"`, `architectures: ["LlamaForCausalLM"]`
+- **Przyczyna:** SpinQuant zamienia architekturę na LlamaForCausalLM podczas rotacji
+- **Rozwiązanie:** Ręczna naprawa config.json + usunięcie pól specyficznych dla Llama (`attention_bias`, `head_dim`, `mlp_bias`, `pretraining_tp`, `rope_scaling`)
+- **Efekt:** Nie naprawił generacji — problem leży głębiej
+
+### Problem 2: Bzdury w open generation (root cause)
+- **Objaw:** Zarówno oryginalny model B jak i nowo-kwantyzowany generują bzdury
+- **Root cause:** Brak runtime transformaty Hadamarda na aktywacjach wejściowych `down_proj`
+- **Mechanizm:**
+  1. SpinQuant aplikuje Hadamard do wag `down_proj` (wygładzenie rozkładu → lepsze dla kwantyzatora)
+  2. W runtime potrzebna jest odwrotna transformata na aktywacjach: `down_proj(Had(x))` zamiast `down_proj(x)`
+  3. gptqmodel (i inne standardowe silniki) nie obsługują tego runtime hooka
+  4. Bez Hadamarda: wagi `down_proj` nie są wygładzone → 2-bit GPTQ wprowadza za duże błędy
+- **Diagnostyka:** Inspekcja R.bin, analiza kodu SpinQuant (`rotation_utils.py`, `hadamard_utils.py`), porównanie logitów FP16 vs quantized
+
+### Problem 3: Próba re-kwantyzacji bez Hadamarda
+- **Hipoteza:** Jeśli pominiemy Hadamard na `down_proj` (żeby nie potrzebować runtime hooka), GPTQ da akceptowalne wyniki
+- **Wynik:** Entropia wzrosła z 1.61 (FP16) do 6.12 (2-bit) — model praktycznie losowy
+- **Wniosek:** Hadamard na `down_proj` jest krytyczny dla jakości 2-bit kwantyzacji. Bez niego wagi mają zbyt duże outliery.
+
+### Problem 4: Infrastruktura
+- gptqmodel wymagał ręcznego patcha `torch/utils/cpp_extension.py` (CUDA 13.0 vs PyTorch cu128 mismatch)
+- CulturaX PL — gated dataset (403), zamieniony na Polish Wikipedia
+- Dysk H200 (79GB) prawie pełny — usunięto hf_cache (21GB) w trakcie kwantyzacji
+
+---
+
+## 4. Wnioski dla paperu
+
+### Główny wniosek
+**SpinQuant + GPTQ 2-bit jest niekompletnym pipeline'em dla standardowego inference.** SpinQuant wymaga runtime transformaty Hadamarda na aktywacjach MLP (`down_proj` input), której żaden standardowy silnik inferencyjny (gptqmodel, AutoGPTQ, vLLM, llama.cpp) nie obsługuje. Bez niej kwantyzacja 2-bit wprowadza destrukcyjne błędy.
+
+### Szczegółowe wnioski
+
+1. **MC eval jest mylący.** Oryginalny model B uzyskał 62.88% na DYK (multiple-choice), mimo że generował kompletne bzdury. MC eval nie testuje spójności generacji — wystarczy, że poprawna odpowiedź ma minimalnie wyższe prawdopodobieństwo niż inne opcje.
+
+2. **SpinQuant R1+R2 bez Hadamarda to za mało.** Rotacja R1 (globalna) i R2 (per-head) poprawiają rozkład wag, ale `down_proj` warstw MLP pozostaje problematyczny. Hadamard na `down_proj` jest trzecim, krytycznym komponentem SmootQuant-style wygładzania.
+
+3. **Entropia jako metryka diagnostyczna.** Porównanie entropii logitów (FP16: 1.61 vs Q2: 6.12–9.00) jest szybkim i jednoznacznym testem jakości kwantyzacji — znacznie bardziej informatywnym niż MC accuracy.
+
+4. **2-bit GPTQ wymaga albo:**
+   - Pełnego SpinQuant pipeline z custom runtime (Hadamard hooks) — niepraktyczne dla deploymentu
+   - Metod natywnie obsługujących 2-bit: QuIP#, AQLM, HQQ z odpowiednimi kernelami
+   - Wyższej precyzji (3-bit, 4-bit) gdzie Hadamard smoothing jest mniej krytyczny
+
+5. **Config fix (llama→mistral) jest konieczny ale niewystarczający.** Poprawia ładowanie modelu i attention implementation, ale nie naprawia fundamentalnego problemu z brakiem Hadamarda.
+
+### Rekomendacja
+Dla Bielik-11B-v2.3-Instruct w 2-bit: rozważyć **QuIP#** (University of Cornell) lub **AQLM** (Egiazarian et al.) — metody zaprojektowane od podstaw pod ekstremalnie niską precyzję, z własnymi kernelami inferencyjnymi obsługującymi transformaty na aktywacjach.
+
+---
+
+*Wygenerowano: 2026-02-22 | GPU: NVIDIA H200 80GB | Software: gptqmodel 5.7.0, transformers 5.2.0*

variant_b/report/variant_b_summary_short.md

@@ -0,0 +1,79 @@
+# Variant B - SpinQuant + GPTQModel na Bielik-11B
+
+## Pipeline
+
+| Step | Opis | Status | Czas |
+|------|------|--------|------|
+| 1-2  | Setup, patche, Mistral->Llama compat | DONE | ~15 min |
+| 3    | Cayley SGD rotation optimization (100 krokow) | DONE | 3h 15m |
+| 4a   | Aplikacja rotacji do wag modelu | DONE | ~5 min |
+| 4b   | GPTQ 2-bit quantization (gptqmodel + Triton V2) | DONE | ~15 min |
+| 5    | Evaluation (lm-evaluation-harness polish3) | PARTIAL | przerwany |
+
+## Step 3 - Rotation Optimization (Cayley SGD)
+
+- Parametry: lr=1.5, cosine schedule, bf16, groupsize=128, 100 steps
+- Kalibracja: 512 samples z polskiej Wikipedii (nie ang. WikiText-2)
+- Loss curve: 21.60 -> 6.79 (step 100)
+- Grad norm: 176.0 -> 1.77 (stabilny od step ~60)
+- Czas per step: ~117s
+- Czas total: ~3h 15min
+- R.bin: 70,398,323 bytes (67 MB)
+- GPU VRAM: ~101 GB / 144 GB
+
+## Step 4 - Quantization
+
+### 4a - Rotation Application
+- rotate_model(model, args) - jedyne API potrzebne
+- Bug: fuse_ln_fcs nie istnieje -> usuniete (rotate_model robi wszystko wewnetrznie)
+- Bug: device_map=cpu -> zmienione na auto (rotate_model uzywa .cuda())
+
+### 4b - GPTQ 2-bit Packing
+- Finalna metoda: gptqmodel + Triton V2 kernel (GPU-accelerated)
+- Parametry: bits=2, group_size=128, desc_act=False, sym=True
+- Kalibracja: 128 samples z pl_wiki_calib.jsonl, max_length=2048
+- Czas quantization: ~10 min (50 layers)
+- Czas packing: ~5 min
+- Model size: 3.3 GB (vs ~22 GB FP16) -> 6.7x kompresja
+
+### Bugi i rozwiazania (Step 4b)
+1. auto-gptq + desc_act=True: CPU packing trwal 2+ godziny, nigdy nie skonczyl
+2. auto-gptq + desc_act=False: nadal wolny (~1h+) bez CUDA extensions
+3. auto-gptq CUDA compilation: deprecated PyTorch API (vec.type()) niekompatybilny z torch 2.6+
+4. gptqmodel: wymagal torch >= 2.7.1 -> upgrade do 2.10.0+cu128
+5. transformers 5.2.0: usuniela no_init_weights -> downgrade do 4.48.3
+6. Smoke test: garbage output ale benchmark MC daje 62.88%
+
+## Step 5 - Evaluation (PARTIAL)
+
+### Wynik DYK MC (jedyny ukonczony task)
+| Task | Metric | Value | Stderr |
+|------|--------|-------|--------|
+| polish_dyk_multiple_choice | acc | 0.6288 | +/-0.0151 |
+| polish_dyk_multiple_choice | acc_norm | 0.6288 | +/-0.0151 |
+| polish_dyk_multiple_choice | f1 | 0.2267 | N/A |
+
+### Status
+- Model NIE jest zepsuty - accuracy 62.88% >> 30% threshold
+- Pelny eval polish_mc przerwany (belebele dataset_name fix zrobiony ale nie dokonczony)
+- Eval polish_generate_few i remaining tasks - nie uruchomiony
+
+## Srodowisko (finalne)
+- GPU: NVIDIA H200 (144 GB VRAM), vast.ai
+- torch: 2.10.0+cu128
+- transformers: 4.43.4
+- gptqmodel: uzyty do quantization, potem uninstalled
+- auto-gptq: 0.7.1 (no CUDA ext) - do eval inference
+- lm-evaluation-harness: polish3 branch
+
+## Artefakty
+- Model: Jakubrd4/bielik-q2-variant-b (HuggingFace)
+- R.bin: w repo modelu (rotations/R.bin, 67 MB)
+- Logi: w repo docs (variant_b/)
+
+## Wnioski
+- SpinQuant rotacje skutecznie obnizily loss (21.6->6.79)
+- 2-bit GPTQ daje 6.7x kompresje (22 GB -> 3.3 GB)
+- DYK MC accuracy 62.88% - obiecujacy wynik, ale pelny benchmark nieukonczony
+- Baseline IQ2_XXS: 61.34% - na razie Variant B prowadzi o ~1.5pp na jednym tasku
+- Potrzebny pelny eval zeby potwierdzic ogolny wynik