---
license: apache-2.0
base_model: Qwen/Qwen3.5-0.8B
library_name: transformers
language:
  - ca
  - en
tags:
  - catalan
  - conversational
  - sft
  - lora-merged
  - projecte-aina
  - bsc-lt
  - alia
datasets:
  - BSC-LT/ALIA-2606-SFT
  - projecte-aina/MentorCA
---

# Gala

**Gala** is a Catalan-focused conversational model, fine-tuned from [Qwen/Qwen3.5-0.8B](https://huggingface.co/Qwen/Qwen3.5-0.8B) on Catalan chat data from [BSC-LT](https://huggingface.co/BSC-LT) and [projecte-aina](https://huggingface.co/projecte-aina).

- **Base:** Qwen/Qwen3.5-0.8B (chat variant, Apache-2.0, multimodal — the vision tower is untouched; all 192 trained modules are language-model projections)
- **Method:** LoRA SFT (r=16, α=32, dropout 0.05) on the language model only, merged into full weights before publish
- **Data:** 43,478 conversations — 35k Catalan slice of [BSC-LT/ALIA-2606-SFT](https://huggingface.co/datasets/BSC-LT/ALIA-2606-SFT) (the SFT mix behind BSC's ALIA-40b-instruct), 7,122 [projecte-aina/MentorCA](https://huggingface.co/datasets/projecte-aina/MentorCA) train rows converted to chat turns, 1,750 English anchor (~4%); assistant-only loss, packing at 2,048 tokens, 2 epochs (350 steps)
- **Hardware:** 21 h 22 m on a single T4 (fp16, `flash-linear-attention` kernel active)
- **Final train loss:** 1.496 → 1.074 · **eval loss (MentorCA validation):** 1.339 · token accuracy 72.6%

## Catalan evaluation (Gala vs base Qwen3.5-0.8B)

Generation-based evaluation, letter-parsing for multiple choice, sampled decoding (temp 0.7, top_p 0.9, repetition penalty 1.1). Full JSON: private `luispoveda93/Gala-project-report/eval_all_v4.json`.

| Benchmark | Gala | Base | Δ |
|---|---|---|---|
| [IFEval_ca](https://huggingface.co/datasets/projecte-aina/IFEval_ca) (strict acc, 150-prompt verifier subset) | **41.3%** | 32.7% | **+8.7** |
| [hhh_alignment_ca](https://huggingface.co/datasets/projecte-aina/hhh_alignment_ca) (overall) | **15.8%** | 10.4% | **+5.4** |
| [tecla](https://huggingface.co/datasets/projecte-aina/tecla) (4-way news classification, n=300) | **44.3%** | 25.3% | **+19.0** |
| [COPA-ca](https://huggingface.co/datasets/projecte-aina/COPA-ca) (n=500) | **64.0%** | 54.0% | **+10.0** |
| [arc_ca](https://huggingface.co/datasets/projecte-aina/arc_ca) Easy (n=300) | **64.7%** | 59.3% | **+5.3** |
| arc_ca Challenge (n=300) | **55.7%** | 54.0% | +1.7 |
| [mgsm_ca](https://huggingface.co/datasets/projecte-aina/mgsm_ca) (n=250) | **23.6%** | 22.4% | +1.2 |
| [EQ-bench_ca](https://huggingface.co/datasets/BSC-LT/EQ-bench_ca) (MAE 0–10, lower better, 160/167 parsed) | **2.96** | 3.35 | −0.39 |
| [CaBBQ](https://huggingface.co/datasets/BSC-LT/CaBBQ) (overall, n=400) | 41.5% | **45.0%** | −3.5 |
| [xquad-ca](https://huggingface.co/datasets/projecte-aina/xquad-ca) (F1, n=300) | 37.9 | **56.1** | **−18.2** |

**Summary:** Gala beats the base model on both core-gate metrics (instruction following and alignment) and on 7 of 10 reported measurements, with the largest gains in Catalan cultural knowledge (Tecla), commonsense (COPA) and instruction following (IFEval_ca).

## Limitations (honest)

- **Fluency ceiling:** long-form Catalan output still contains grammatical errors, invented words and hallucinated content — a limitation of the 0.8B scale and the SFT-only recipe, not fixable by more of the same data. Short conversational turns are noticeably more reliable than long generations.
- **Extractive QA regression:** on XQuAD-ca the base model's F1 is higher; Gala's chattier style hurts span-extraction tasks.
- **CaBBQ** is slightly lower than base; MC bias-probe agreement at 0.8B is near chance for both models and should be read cautiously.
- Both models were evaluated with the same harness; small-model MC letter-parsing adds noise to all absolute numbers.
- Multimodal (vision) behaviour is inherited from the base model and was **not** fine-tuned or evaluated.

## Intended use

Catalan-first conversational assistant for experimentation and demos. Not recommended for production factual use without guardrails.

## Training data licenses

ALIA-2606-SFT (CC-BY-4.0) · MentorCA (CC-BY-SA-4.0) · English anchor from ALIA-2606-SFT. [InstruCAT](https://huggingface.co/datasets/projecte-aina/InstruCAT) (CC-BY-NC-ND) was deliberately excluded.

**Decision-reasoning report:** private repo [luispoveda93/Gala-project-report](https://huggingface.co/luispoveda93/Gala-project-report) · **Training dashboard:** https://huggingface.co/spaces/luispoveda93/gala-catalan-sft-trackio