mistral-12b-sft / README.md
kosmylo1992's picture
Update README.md
d0620d4 verified
metadata
language:
  - en
license: apache-2.0
tags:
  - text-generation
  - causal-lm
  - instruction-tuning
  - supervised-fine-tuning
  - synthetic-qa
  - lora
  - axolotl
  - deepspeed
  - transformers
  - mistral
  - nemo
  - eu-hpc
datasets:
  - axolotl_deduplicated_synthetic_qa
metrics:
  - loss
library_name: transformers
framework: pytorch
base_model: mistralai/Mistral-Nemo-Instruct-2407
model_name: mistral-12b-sft
pipeline_tag: text-generation
task_categories:
  - text-generation
  - instruction-following
model_type: AutoModelForCausalLM
inference:
  parameters:
    max_new_tokens: 512
    temperature: 0.7
    top_p: 0.9
trained_on:
  - Leonardo EuroHPC
description: >-
  Supervised fine-tuning (SFT) of Mistral 12B Nemo Instruct on synthetic QA data
  using LoRA with Axolotl and DeepSpeed. Improves conversational reasoning and
  factual accuracy.

Mistral 12B — SFT (Supervised Fine-Tuning on Synthetic QA)

Model type: Causal Language Model
Base model: mistralai/Mistral-Nemo-Instruct-2407
License: Apache 2.0
Framework: Axolotl


Overview

mistral-12b-sft is a supervised fine-tuned variant of Mistral-12B trained on high-quality synthetic QA data.
This SFT phase enhances instruction following, factual reasoning, and conversational ability while maintaining model efficiency via 8-bit LoRA adapters.

Training was conducted on Leonardo EuroHPC.


Training Setup

Objective: Supervised fine-tuning (instruction-following QA)
Adapter: LoRA + 8-bit base
Precision: bfloat16
Hardware: 8 × 2 × A100 64 GB
Framework: Axolotl + DeepSpeed + PyTorch 2.5.1 + CUDA 12.1
Runtime: ~6 h
Validation: 30 %


Dataset

Dataset Type Description
axolotl_deduplicated_synthetic_qa.jsonl alpaca_chat.load_qa Synthetic instruction–response pairs for QA and chat fine-tuning

Hyperparameters

Parameter Value
Sequence length 2048
Micro batch size 2
Gradient accumulation 2
Epochs 1
Learning rate 0.0002
LR scheduler cosine
Optimizer AdamW (8-bit)
Warmup steps 10
Weight decay 0.0
LoRA rank (r) 16
LoRA alpha 32
LoRA dropout 0.05
LoRA targets q_proj, k_proj, v_proj, o_proj
Gradient checkpointing
Flash attention
Auto-resume
Loss watchdog threshold 5.0, patience 3

Tokenizer

Tokenizer type: AutoTokenizer
Pad token: <|end_of_text|>