salamandra-7b GGUF ladder (imatrix es) + KLD/PPL/Top-1 fidelity metrics + charts

.gitattributes

@@ -33,3 +33,9 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 *.zip filter=lfs diff=lfs merge=lfs -text
 *.zst filter=lfs diff=lfs merge=lfs -text
 *tfevents* filter=lfs diff=lfs merge=lfs -text
+salamandra-7b-IQ4_XS.gguf filter=lfs diff=lfs merge=lfs -text
+salamandra-7b-Q3_K_M.gguf filter=lfs diff=lfs merge=lfs -text
+salamandra-7b-Q4_K_M.gguf filter=lfs diff=lfs merge=lfs -text
+salamandra-7b-Q5_K_M.gguf filter=lfs diff=lfs merge=lfs -text
+salamandra-7b-Q6_K.gguf filter=lfs diff=lfs merge=lfs -text
+salamandra-7b-Q8_0.gguf filter=lfs diff=lfs merge=lfs -text

Modelfile.4096

@@ -0,0 +1,6 @@
+FROM ./salamandra-7b-Q4_K_M.gguf
+PARAMETER num_ctx 4096
+PARAMETER temperature 0.7
+PARAMETER top_p 0.95
+PARAMETER top_k 40
+PARAMETER repeat_penalty 1.1

Modelfile.8192

@@ -0,0 +1,6 @@
+FROM ./salamandra-7b-Q4_K_M.gguf
+PARAMETER num_ctx 8192
+PARAMETER temperature 0.7
+PARAMETER top_p 0.95
+PARAMETER top_k 40
+PARAMETER repeat_penalty 1.1

README.md

@@ -0,0 +1,135 @@
+---
+license: apache-2.0
+base_model: BSC-LT/salamandra-7b-instruct-2606
+base_model_relation: quantized
+language:
+  - es
+  - ca
+  - gl
+  - eu
+  - en
+pipeline_tag: text-generation
+tags:
+  - gguf
+  - llama.cpp
+  - quantized
+  - imatrix
+  - spanish
+  - salamandra
+  - bsc
+library_name: gguf
+---
+
+<div align="center">
+
+```
+╔══════════════════════════════════════════════════════════════╗
+║   S A L A M A N D R A · 7 B · I N S T R U C T · 2 6 0 6       ║
+║   ───────────────────────────────────────────────────────    ║
+║   GGUF quant ladder · imatrix (calibrado en español)          ║
+║   Q3 ▸ IQ4 ▸ Q4 ▸ Q5 ▸ Q6 ▸ Q8   ·   fidelidad medida (KLD)   ║
+╚══════════════════════════════════════════════════════════════╝
+```
+
+# Salamandra-7B-Instruct-2606 · GGUF
+
+**Cuantizaciones GGUF con imatrix del modelo soberano español/europeo de BSC-LT.**
+Escalera completa Q3→Q8 · matriz de importancia calibrada **en español** · fidelidad objetiva **medida contra el BF16** (KLD + PPL + Top-1).
+
+</div>
+
+> **Qué es esto**: [BSC-LT/salamandra-7b-instruct-2606](https://huggingface.co/BSC-LT/salamandra-7b-instruct-2606) —
+> el modelo multilingüe (ES · CA · GL · EU · EN…) del Barcelona Supercomputing Center— empaquetado en GGUF para
+> correr en local con `llama.cpp` / Ollama. **No es un modelo de código**: su valor es lingüístico y de soberanía,
+> así que aquí lo importante es que **la cuantización no rompa el modelo** — y eso se demuestra con la tabla de fidelidad de abajo, no con un benchmark de programación.
+
+## ✅ Qué archivo elegir
+
+| Si quieres… | Usa | Por qué |
+|---|---|---|
+| **Lo mejor equilibrado** | **Q4_K_M** (4.5 GB) | el estándar seguro — cabe en 8 GB, buena fidelidad |
+| Máxima calidad práctica | **Q6_K** (6.0 GB) | KLD 0.004, casi indistinguible del original |
+| Fidelidad de archivo | **Q8_0** (7.7 GB) | prácticamente sin pérdida (Top-1 98.9%) |
+| Lo más pequeño usable | **Q3_K_M** (3.8 GB) | para RAM muy justa; algo de deriva |
+| Compacto con imatrix | **IQ4_XS** (4.2 GB) | más pequeño que Q4_K_M, fidelidad parecida |
+
+## 📦 Archivos
+
+| Quant | Bits (aprox) | Tamaño |
+|---|---|---|
+| `salamandra-7b-Q3_K_M.gguf` | ~3.9 | 3.8 GB |
+| `salamandra-7b-IQ4_XS.gguf` | ~4.25 | 4.2 GB |
+| `salamandra-7b-Q4_K_M.gguf` | ~4.8 | 4.5 GB |
+| `salamandra-7b-Q5_K_M.gguf` | ~5.6 | 5.2 GB |
+| `salamandra-7b-Q6_K.gguf` | ~6.5 | 6.0 GB |
+| `salamandra-7b-Q8_0.gguf` | ~8.5 | 7.7 GB |
+
+## 📊 Fidelidad medida (vs BF16 · corpus español)
+
+**KLD** (divergencia Kullback–Leibler) mide cuánto se aleja la distribución de salida de cada cuantización respecto al modelo original en BF16 — es el patrón oro de fidelidad de cuantización, más fino que la perplejidad sola. **Top-1** = con qué frecuencia el token más probable coincide con el del BF16.
+
+| Quant | Tamaño | PPL (es) | PPL Δ vs BF16 | **KLD media** | Top-1 vs BF16 |
+|---|---|---|---|---|---|
+| BF16 (ref) | 14.8 GB | 7.437 | — | 0 | 100% |
+| Q8_0 | 7.7 GB | 7.437 | **+0.0002** | **0.00043** | **98.9%** |
+| Q6_K | 6.0 GB | 7.460 | +0.023 | 0.00376 | 96.8% |
+| Q5_K_M | 5.2 GB | 7.485 | +0.048 | 0.00865 | 95.3% |
+| Q4_K_M | 4.5 GB | 7.611 | +0.174 | 0.02808 | 91.9% |
+| IQ4_XS | 4.2 GB | 7.705 | +0.268 | 0.03421 | 91.3% |
+| Q3_K_M | 3.8 GB | 8.117 | +0.680 | 0.08682 | 86.1% |
+
+## 📈 Gráficas
+
+![fidelidad vs tamaño](metrics/chart-quality-vs-size.png)
+![PPL delta](metrics/chart-ppl-delta.png)
+![Top-1 match](metrics/chart-top1-match.png)
+
+## 🧮 ¿Me cabe? (RAM aproximada, contexto 8K)
+
+| Quant | RAM mínima |
+|---|---|
+| Q3_K_M / IQ4_XS | ~6 GB |
+| Q4_K_M / Q5_K_M | ~7–8 GB |
+| Q6_K / Q8_0 | ~8–10 GB |
+
+## 🚀 Cómo ejecutarlo
+
+**Ollama** (usa la plantilla ChatML que ya trae el GGUF):
+```bash
+ollama create salamandra7b -f Modelfile.8192
+ollama run salamandra7b "Explica la fotosíntesis en una frase."
+```
+**llama.cpp**:
+```bash
+llama-cli -m salamandra-7b-Q4_K_M.gguf -ngl 99 -c 8192 \
+  -p "<|im_start|>user\n¿Cuál es la capital de Galicia?<|im_end|>\n<|im_start|>assistant\n"
+```
+Muestreo recomendado: `temp 0.7 · top_p 0.95 · top_k 40`. Contexto nativo **8192**.
+
+## ⚠️ Bueno saber
+
+- **Es un modelo lingüístico, no de código.** Brilla en español y lenguas cooficiales (catalán, gallego, euskera) y multilingüe; no esperes rendimiento de un modelo especializado en programación.
+- **Gate de calidad = fidelidad, no SWE.** Para una cuantización de un modelo así, lo que importa es que se parezca al original (KLD arriba), no un test de coding que sería ~0 por diseño del modelo base.
+- **imatrix en español**: la matriz de importancia se calibró con prosa española, no inglesa — evita la deriva que introduce calibrar en el idioma equivocado.
+
+## 🔁 Procedencia y reproducibilidad
+
+- **Base**: `BSC-LT/salamandra-7b-instruct-2606` · arquitectura Llama · licencia Apache-2.0.
+- **imatrix**: corpus en español, `-c 512`, sobre el BF16.
+- **Ladder**: `llama-quantize --imatrix` para cada tier. Magic verificado en cada `.gguf`.
+- **Fidelidad**: `llama-perplexity --kl-divergence` contra base BF16 (`kld.dat`) + PPL en español.
+- Comandos exactos en `scripts/reproduce.sh`; métricas en `metrics/quant-summary-with-kld.{json,csv}`; checksums en `reports/artifact-sha256sums.txt`.
+
+## Evaluation methodology
+
+- **Métrica**: KLD (mean/max) + Top-1 match + PPL, cada cuantización vs la referencia BF16, sobre un corpus **español**. Sin gate SWE — no aplica a un modelo lingüístico no-coder (sería 0 por diseño, no por la cuantización).
+- **Herramientas**: `llama.cpp` (`llama-imatrix`, `llama-quantize`, `llama-perplexity --kl-divergence`).
+- **Fecha**: 2026-07. Caveat: PPL/KLD sobre un corpus acotado; comparativa relativa entre tiers, no un ranking absoluto.
+
+## 🗒️ Changelog
+
+- **2026-07** — release inicial: escalera Q3→Q8 + IQ4_XS, imatrix español, tabla KLD/PPL/Top-1 + gráficas.
+
+---
+
+Cuantizado por **KikoCis**. Modelo base © Barcelona Supercomputing Center (BSC-LT), Apache-2.0. Todo el mérito del modelo es de BSC-LT; aquí solo se empaqueta y se mide la fidelidad de las cuantizaciones.

metrics/quant-summary-with-kld.csv

@@ -0,0 +1,7 @@
+name,size_bytes,size_gb_decimal,kld_nats_mean,kld_nats_max,kld_nats_p95,kld_nats_p50,top1_match_rate,ppl_c512,ppl_delta_vs_bf16
+Q3_K_M,4047942752,4.05,0.08682,9.011917,0.728921,0.043465,86.063,8.1171,0.6803
+IQ4_XS,4458263648,4.46,0.03421,5.305482,0.287807,0.017275,91.255,7.7047,0.2679
+Q4_K_M,4850562144,4.85,0.028081,21.190672,0.22577,0.014207,91.93,7.6111,0.1743
+Q5_K_M,5591905376,5.59,0.008646,1.567585,0.069664,0.00456,95.298,7.4848,0.048
+Q6_K,6379582560,6.38,0.003758,0.422751,0.029011,0.002122,96.811,7.4601,0.0233
+Q8_0,8260858976,8.26,0.000433,0.519766,0.003259,0.000223,98.88,7.437,0.0002

metrics/quant-summary-with-kld.json

@@ -0,0 +1,80 @@
+{
+  "model": "salamandra-7b-instruct-2606",
+  "ref": "BF16",
+  "ref_ppl": 7.4368,
+  "corpus": "es",
+  "rows": [
+    {
+      "name": "Q3_K_M",
+      "size_bytes": 4047942752,
+      "size_gb_decimal": 4.05,
+      "kld_nats_mean": 0.08682,
+      "kld_nats_max": 9.011917,
+      "kld_nats_p95": 0.728921,
+      "kld_nats_p50": 0.043465,
+      "top1_match_rate": 86.063,
+      "ppl_c512": 8.1171,
+      "ppl_delta_vs_bf16": 0.6803
+    },
+    {
+      "name": "IQ4_XS",
+      "size_bytes": 4458263648,
+      "size_gb_decimal": 4.46,
+      "kld_nats_mean": 0.03421,
+      "kld_nats_max": 5.305482,
+      "kld_nats_p95": 0.287807,
+      "kld_nats_p50": 0.017275,
+      "top1_match_rate": 91.255,
+      "ppl_c512": 7.7047,
+      "ppl_delta_vs_bf16": 0.2679
+    },
+    {
+      "name": "Q4_K_M",
+      "size_bytes": 4850562144,
+      "size_gb_decimal": 4.85,
+      "kld_nats_mean": 0.028081,
+      "kld_nats_max": 21.190672,
+      "kld_nats_p95": 0.22577,
+      "kld_nats_p50": 0.014207,
+      "top1_match_rate": 91.93,
+      "ppl_c512": 7.6111,
+      "ppl_delta_vs_bf16": 0.1743
+    },
+    {
+      "name": "Q5_K_M",
+      "size_bytes": 5591905376,
+      "size_gb_decimal": 5.59,
+      "kld_nats_mean": 0.008646,
+      "kld_nats_max": 1.567585,
+      "kld_nats_p95": 0.069664,
+      "kld_nats_p50": 0.00456,
+      "top1_match_rate": 95.298,
+      "ppl_c512": 7.4848,
+      "ppl_delta_vs_bf16": 0.048
+    },
+    {
+      "name": "Q6_K",
+      "size_bytes": 6379582560,
+      "size_gb_decimal": 6.38,
+      "kld_nats_mean": 0.003758,
+      "kld_nats_max": 0.422751,
+      "kld_nats_p95": 0.029011,
+      "kld_nats_p50": 0.002122,
+      "top1_match_rate": 96.811,
+      "ppl_c512": 7.4601,
+      "ppl_delta_vs_bf16": 0.0233
+    },
+    {
+      "name": "Q8_0",
+      "size_bytes": 8260858976,
+      "size_gb_decimal": 8.26,
+      "kld_nats_mean": 0.000433,
+      "kld_nats_max": 0.519766,
+      "kld_nats_p95": 0.003259,
+      "kld_nats_p50": 0.000223,
+      "top1_match_rate": 98.88,
+      "ppl_c512": 7.437,
+      "ppl_delta_vs_bf16": 0.0002
+    }
+  ]
+}

reports/artifact-sha256sums.txt

@@ -0,0 +1,8 @@
+7a775454cdb1be9fc3ac4c4c51ef4691ba474867d283b95b268fcac14d853c8a  salamandra-7b-Q3_K_M.gguf
+0643c080734e33486f0f10bd721b502e33b1eeaf5a0a1dd754c7466dd36f6c2d  salamandra-7b-Q4_K_M.gguf
+0d9ba4318e8c73c46d729ac62b9a65897ad57a22918545b39b64cb9d7db324e5  salamandra-7b-Q5_K_M.gguf
+8837ea5c00cbb6bbb3ae3775a5cdc5369c32e1aa7b59a961a6c2a1ecfd87f8b0  salamandra-7b-Q6_K.gguf
+f49e470cec70107b2d173fe05ccfad630d27876758328ce90762626e9eb96bd8  salamandra-7b-Q8_0.gguf
+61c60dd48df883b8ad3ae05ed0f1fbf4327f9e07879f1d54ac3cf2bc959456ae  salamandra-7b-IQ4_XS.gguf
+efe8ce6a43d37f787b5de5bc46e9bd9b21eb81842b97738f84f00f331e99bc56  metrics/quant-summary-with-kld.json
+3954710b67300253ce5c8ecb5491b30a32d5df18b82170b4b80f3618d70d6406  metrics/quant-summary-with-kld.csv

salamandra-7b-IQ4_XS.gguf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:61c60dd48df883b8ad3ae05ed0f1fbf4327f9e07879f1d54ac3cf2bc959456ae
+size 4458263648

salamandra-7b-Q3_K_M.gguf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:7a775454cdb1be9fc3ac4c4c51ef4691ba474867d283b95b268fcac14d853c8a
+size 4047942752

salamandra-7b-Q4_K_M.gguf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0643c080734e33486f0f10bd721b502e33b1eeaf5a0a1dd754c7466dd36f6c2d
+size 4850562144

salamandra-7b-Q5_K_M.gguf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0d9ba4318e8c73c46d729ac62b9a65897ad57a22918545b39b64cb9d7db324e5
+size 5591905376

salamandra-7b-Q6_K.gguf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:8837ea5c00cbb6bbb3ae3775a5cdc5369c32e1aa7b59a961a6c2a1ecfd87f8b0
+size 6379582560

salamandra-7b-Q8_0.gguf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f49e470cec70107b2d173fe05ccfad630d27876758328ce90762626e9eb96bd8
+size 8260858976

scripts/kld_sweep.py

@@ -0,0 +1,61 @@
+#!/usr/bin/env python3
+"""KLD + PPL + Top-1 sweep for the salamandra-7b quant ladder vs the BF16 reference.
+Waits for the ladder, builds the KLD base from BF16, then per-quant measures KLD (mean/p50/p95/max,
+nats) + Top-1 match rate + PPL (Spanish corpus). Writes metrics/quant-summary-with-kld.{json,csv}.
+"""
+import os, re, json, csv, time, subprocess, glob
+D = "/Users/kikocisneros/coco_ppl/aq_salamandra7b"
+REF = f"{D}/salamandra-7b-BF16.gguf"
+CORPUS = f"{D}/es_prose.txt"          # Spanish (model's language)
+KLDBASE = f"{D}/kld.dat"
+METR = f"{D}/metrics"; os.makedirs(METR, exist_ok=True)
+CTX = "512"; LADDER = ["Q3_K_M", "IQ4_XS", "Q4_K_M", "Q5_K_M", "Q6_K", "Q8_0"]
+
+def log(m): print(f"[kld {time.strftime('%H:%M:%S')}] {m}", flush=True)
+def sh(cmd):
+    p = subprocess.run(cmd, stdout=subprocess.PIPE, stderr=subprocess.STDOUT, text=True)
+    return p.stdout
+
+def gnum(pat, text, d=float("nan")):
+    m = re.search(pat, text)
+    return float(m.group(1)) if m else d
+
+# wait for ladder
+while not os.path.exists(f"{D}/salamandra-7b-Q8_0.gguf") or subprocess.run(["pgrep","-f","llama-quantize"],capture_output=True).returncode==0:
+    time.sleep(8)
+log("ladder present; building KLD base from BF16")
+sh(["llama-perplexity","-m",REF,"-f",CORPUS,"-c",CTX,"--kl-divergence-base",KLDBASE,"-ngl","99"])
+log("KLD base built")
+
+rows = []
+for Q in LADDER:
+    g = f"{D}/salamandra-7b-{Q}.gguf"
+    if not os.path.exists(g): log(f"skip {Q} (missing)"); continue
+    size = os.path.getsize(g)
+    log(f"KLD {Q}")
+    kld = sh(["llama-perplexity","-m",g,"--kl-divergence-base",KLDBASE,"--kl-divergence","-ngl","99"])
+    log(f"PPL {Q}")
+    ppl = sh(["llama-perplexity","-m",g,"-f",CORPUS,"-c",CTX,"-ngl","99"])
+    row = {
+        "name": Q, "size_bytes": size, "size_gb_decimal": round(size/1e9, 2),
+        "kld_nats_mean": gnum(r"[Mm]ean\s+KLD\s*[:=]\s*([0-9.eE-]+)", kld),
+        "kld_nats_max":  gnum(r"[Mm]ax(?:imum)?\s+KLD\s*[:=]\s*([0-9.eE-]+)", kld),
+        "kld_nats_p95":  gnum(r"99\.0%\s+KLD\s*[:=]\s*([0-9.eE-]+)", kld),
+        "kld_nats_p50":  gnum(r"[Mm]edian\s+KLD\s*[:=]\s*([0-9.eE-]+)", kld),
+        "top1_match_rate": gnum(r"[Ss]ame top p:\s*([0-9.]+)", kld) or gnum(r"top.?1.*?([0-9.]+)%", kld),
+        "ppl_c512": gnum(r"PPL\s*[:=]\s*([0-9.]+)", ppl) or gnum(r"perplexity:\s*([0-9.]+)", ppl) or gnum(r"Final estimate:\s*PPL\s*=\s*([0-9.]+)", ppl),
+    }
+    rows.append(row)
+    log(f"  {Q}: KLD_mean={row['kld_nats_mean']} top1={row['top1_match_rate']} ppl={row['ppl_c512']}")
+    json.dump({"model":"salamandra-7b-instruct-2606","ref":"BF16","corpus":"es","rows":rows},
+              open(f"{METR}/quant-summary-with-kld.json","w"), indent=2)
+
+# ppl delta vs bf16 ref (compute ref ppl once)
+refppl = sh(["llama-perplexity","-m",REF,"-f",CORPUS,"-c",CTX,"-ngl","99"])
+rp = gnum(r"Final estimate:\s*PPL\s*=\s*([0-9.]+)", refppl) or gnum(r"PPL\s*[:=]\s*([0-9.]+)", refppl)
+for r in rows: r["ppl_delta_vs_bf16"] = round(r["ppl_c512"]-rp,4) if r["ppl_c512"]==r["ppl_c512"] and rp==rp else None
+json.dump({"model":"salamandra-7b-instruct-2606","ref":"BF16","ref_ppl":rp,"corpus":"es","rows":rows},
+          open(f"{METR}/quant-summary-with-kld.json","w"), indent=2)
+with open(f"{METR}/quant-summary-with-kld.csv","w",newline="") as f:
+    w = csv.DictWriter(f, fieldnames=list(rows[0].keys())); w.writeheader(); w.writerows(rows)
+log("KLD_SWEEP_DONE")

scripts/make_charts.py

@@ -0,0 +1,27 @@
+import json, matplotlib
+matplotlib.use("Agg")
+import matplotlib.pyplot as plt
+D="/Users/kikocisneros/coco_ppl/aq_salamandra7b"
+d=json.load(open(f"{D}/metrics/quant-summary-with-kld.json"))
+rows=d["rows"]; names=[r["name"] for r in rows]
+gb=[r["size_gb_decimal"] for r in rows]; kld=[r["kld_nats_mean"] for r in rows]
+ppld=[r["ppl_delta_vs_bf16"] for r in rows]; top1=[r["top1_match_rate"] for r in rows]
+C="#3b5bdb"; G="#0ca678"; O="#e8590c"
+def style(ax,t,xl,yl): ax.set_title(t,fontweight="bold"); ax.set_xlabel(xl); ax.set_ylabel(yl); ax.grid(alpha=.25)
+# 1. quality vs size (KLD vs GB, log y) — the money chart
+fig,ax=plt.subplots(figsize=(7,4.5))
+ax.plot(gb,kld,"-o",color=C,lw=2,ms=8)
+for x,y,n in zip(gb,kld,names): ax.annotate(n,(x,y),textcoords="offset points",xytext=(6,6),fontsize=9)
+ax.set_yscale("log"); style(ax,"Fidelity vs size — salamandra-7b GGUF","file size (GB)","KLD mean vs BF16 (nats, log) — lower is better")
+fig.tight_layout(); fig.savefig(f"{D}/metrics/chart-quality-vs-size.png",dpi=130); plt.close()
+# 2. ppl delta
+fig,ax=plt.subplots(figsize=(7,4))
+b=ax.bar(names,ppld,color=[G if v<0.1 else (C if v<0.4 else O) for v in ppld])
+for r,v in zip(b,ppld): ax.text(r.get_x()+r.get_width()/2,v,f"+{v:.3f}",ha="center",va="bottom",fontsize=9)
+style(ax,"Perplexity increase vs BF16 (Spanish)","","PPL Δ — lower is better"); fig.tight_layout(); fig.savefig(f"{D}/metrics/chart-ppl-delta.png",dpi=130); plt.close()
+# 3. top-1 match
+fig,ax=plt.subplots(figsize=(7,4))
+b=ax.bar(names,top1,color=G)
+for r,v in zip(b,top1): ax.text(r.get_x()+r.get_width()/2,v,f"{v:.1f}%",ha="center",va="bottom",fontsize=9)
+ax.set_ylim(80,100); style(ax,"Top-1 token agreement with BF16","","% argmax matches — higher is better"); fig.tight_layout(); fig.savefig(f"{D}/metrics/chart-top1-match.png",dpi=130); plt.close()
+print("CHARTS_DONE", [n for n in names])

scripts/reproduce.sh

@@ -0,0 +1,10 @@
+# salamandra-7b-instruct-2606 GGUF — reproduction
+# 1. convert reference
+python llama.cpp/convert_hf_to_gguf.py BSC-LT/salamandra-7b-instruct-2606 --outfile salamandra-7b-BF16.gguf --outtype bf16
+# 2. imatrix — calibrated in SPANISH (the model's language)
+llama-imatrix -m salamandra-7b-BF16.gguf -f imatrix_calib_es.txt -o salamandra-7b.imatrix -c 512 -ngl 99
+# 3. quantize the ladder
+for Q in Q3_K_M IQ4_XS Q4_K_M Q5_K_M Q6_K Q8_0; do
+  llama-quantize --imatrix salamandra-7b.imatrix salamandra-7b-BF16.gguf salamandra-7b-$Q.gguf $Q
+done
+# 4. KLD/PPL/Top-1 sweep vs BF16 (Spanish corpus) -> metrics/quant-summary-with-kld.json