~/sameer-batra
← home

Research

How far can small models go?

Harness vs Model

Complete

Independent experiment

Can a general LLM running locally replace a purpose-built 200M-parameter specialist for knowledge-graph extraction? Turning off one default setting, "thinking", moved relation F1 from 0.33 to 0.51.

EvaluationKnowledge graphsDebuggingRead the case study →

FactAgent-Bench

In progress

Capstone research

Capstone with two teammates: do multi-agent critique and debate pipelines improve factual consistency in 7–8B open-weight models on a single GPU, or is reinforcement learning a better use of the compute?

Multi-agentEvaluationGRPO

Context-length extension

Complete

Research experiment

Small models can beat large ones on a specific task but cannot read long inputs. I tested two context-extension methods, Linear Position Interpolation and YaRN, on Phi-2 and TinyLlama. The simpler one, Linear PI, held up best: with it, Phi-2 perplexity on 8K-token inputs fell from 286 to 15 (lower is better).

Fine-tuningPhi-2YaRNRead the case study →

Education