bhuvan0808/buvn-2.0-sft

What is BUVN-2.0-SFT?

This is the instruction-tuned version of BUVN-2.0, fine-tuned on conversational data from Alpaca (52K instructions) and OpenAssistant (13.9K English conversation pairs).

The model learns to follow the chat template format and generate responses in a conversational style.

Training Details

Setting	Value
Base Model	buvn-2.0 (109.5M params)
SFT Data	52,002 Alpaca + 13,892 OpenAssistant = 65,894 examples
Total Tokens	9,194,803
Training Steps	3,000
Learning Rate	3×10⁻⁵ (20x lower than pre-training)
Batch Size	32 × 4 grad accum = 128 effective
Best Val Loss	1.8337 (PPL 6.3)
Training Time	~25 min on H100 NVL
Precision	bfloat16 + torch.compile

Training Progress

Step     0: val loss 3.19 (ppl 24.4) ← loaded pre-trained weights
Step   200: val loss 1.98 (ppl  7.2) ← rapidly learning chat format
Step   600: val loss 1.87 (ppl  6.5)
Step  1000: val loss 1.85 (ppl  6.4)
Step  2000: val loss 1.84 (ppl  6.3)
Step  3000: val loss 1.83 (ppl  6.3) ← converged

Chat Template

The model was trained with this conversation format:

<|user|>
What is the capital of France?
<|end|>
<|assistant|>
The capital of France is Paris.
<|end|>

Honest Assessment

At 109.5M parameters, the model successfully learns the chat format but has limited instruction-following accuracy. This is a known limitation of small models:

Model Size	Instruction Quality
109.5M (this model)	Learns format; answers often tangential to the question
350M+	Basic Q&A works for simple questions
1B+	Solid instruction following
3B+ (Phi-2, TinyLlama)	Good quality, genuinely useful
7B+ (LLaMA, Mistral)	Strong instruction following

The model needs to be scaled to 350M+ parameters for meaningful instruction following. The SFT pipeline and infrastructure are validated and ready for scaling.

Files

File	Size	Description
buvn_2.0_sft_best.pt	1.31 GB	SFT fine-tuned checkpoint
tokenizer_32k.json	2.2 MB	32K BPE tokenizer
config.json	~200 B	Model hyperparameters
sft_meta.json	~500 B	Training data statistics

The Beuvian Ecosystem

Model	Status	Description
BUVN-2.0	✅ Released	Foundation model (PPL 29.19, beats GPT-2 Small)
BUVN-2.0-SFT (this)	✅ Released	Instruction-tuned version
SRVN	🔜 Planned	Code agent (fine-tuned on code data)
MNI	🔜 Planned	Finance model (market data, SEC filings)

Links

• Base model: bhuvan0808/buvn-2.0
• GitHub: bhuvan0808/beuvian
• Documentation: docs/

Built by Bhuvan | Beuvian AI Ecosystem