eulogik/Bharat-Tiny-LLM-v3

fix remaining compression table numbers (21.6→36.8, 16.7→18.0)

fix compression numbers (21.6→36.8, 16.7→18.0) to match paper/bench; update demo badge to Featherless

update banner: compressed to 429KB

add banner to model card

add banner image (shared across model card, Space, GitHub)

fix: restore v3 safetensors (HEAD was v3.1/v3.2-era conversational variant)

fix: restore v3 generation_config.json (HEAD was v3.1/v3.2-era)

fix: restore v3 tokenizer.json (HEAD was v3.1/v3.2-era)

fix: restore v3 tokenizer_config.json (HEAD was v3.1/v3.2-era)

fix: restore v3 config.json (HEAD was v3.1/v3.2-era)

doc: add v3.3 variant note to body

doc: add v3.3 variant note + gsm8k, math, conversational-ai tags

doc: add v3.3 conversational GGUF, datasets, tags

v3.3: Q4_K_M GGUF

v3.3: 370 curated rows SFT GGUF F16

v3.1: conversational SFT on samvaad+hinglish+corpus

v3.2 GGUF: brahmi-v3.2-q4_k_m.gguf (2.5K curated, r=8)

v3.2 GGUF: brahmi-v3.2-f16.gguf (2.5K curated, r=8)

v3.1: conversational SFT on samvaad+hinglish+corpus

revert to v3: conversational SFT degraded quality (kept as v3.1 GGUFs)

v3.1 GGUF: brahmi-v3.1-q4_k_m.gguf

v3.1 GGUF: brahmi-v3.1-f16.gguf

v3.1: conversational SFT on samvaad+hinglish+corpus

remove extra_special_tokens list (causes TypeError in transformers)

add added_tokens_decoder for slow tokenizer support

add merges.txt for slow tokenizer

add vocab.json for slow tokenizer

SEO/AEO/GEO-optimized model card with badges + rich tags

Add model card with v3 benchmarks

Add bharat-tiny-llm-v3-q4_k_m.gguf for edge deployment

Add bharat-tiny-llm-v3-f16.gguf for edge deployment

v3: Qwen3-1.7B + Devanagari injection + warmup + SFT (GSM8K-Hindi 56%)

v3: Qwen3-1.7B + Brahmi-300 + capability SFT

initial commit