Skip to content

Add Glyphh Ada 1.1 (HDC+FC) to BFCL V4 Leaderboard - #1308

Open
timmetim wants to merge 2 commits into
ShishirPatil:mainfrom
timmetim:glyphh-ada-1.1
Open

Add Glyphh Ada 1.1 (HDC+FC) to BFCL V4 Leaderboard#1308
timmetim wants to merge 2 commits into
ShishirPatil:mainfrom
timmetim:glyphh-ada-1.1

Conversation

@timmetim

Copy link
Copy Markdown

Summary

Adds Glyphh Ada 1.1, a hybrid Hyperdimensional Computing (HDC) + Function Calling model to the BFCL V4 leaderboard.

Architecture

Glyphh Ada uses a two-stage pipeline:

  1. HDC Routing (deterministic, no LLM) — Encodes queries and function signatures into 10,000-dimensional hypervectors, then uses cosine similarity to route to the top-3 candidate functions
  2. Argument Extraction (Claude Haiku 4.5) — Only the HDC-matched functions are sent to the LLM for parameter extraction via native tool_use

This means function selection is deterministic and instant, while the LLM is only used for argument extraction on a filtered candidate set.

Results

Category Accuracy
Overall 74.50%
Non-Live 88.71%
Live 74.32%
Irrelevance 87.56%
Multi-Turn 53.75%
Web Search 79.50%
Memory 87.10%

Total cost: $2.08 | Mean latency: 8.52s

Changes

  • bfcl_eval/model_handler/api_inference/glyphh.pyGlyphhHandler extending ClaudeHandler
  • bfcl_eval/constants/model_config.py — Model config with Haiku pricing ($0.05/$0.40 per 1M tokens)
  • bfcl_eval/constants/supported_models.py — Added glyphh-ada-1.1
  • SUPPORTED_MODELS.md — Added to supported models table
  • result/glyphh-ada-1.1/ — Pre-computed result files for all categories
  • score/ — Evaluation scores generated by eval_runner.py --test-category all

All results verified locally using eval_runner.py.

Christopher Timmerman and others added 2 commits March 10, 2026 16:02
Adds GlyphhHDCHandler for decoding pre-computed HDC model results.
Handles JSON-stringified parameters in decode_execute.
Registers glyphh-hdc-v1 and glyphh-ada-1.1 model configs.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Add Glyphh Ada 1.1, a hybrid Hyperdimensional Computing (HDC) + Function
Calling model. Uses HDC for deterministic function routing (cosine similarity
in 10,000-dimensional space) and Claude Haiku 4.5 for argument extraction.

Overall accuracy: 74.50%
- Non-Live: 88.71%
- Live: 74.32%
- Irrelevance: 87.56%
- Multi-Turn: 53.75%
- Web Search: 79.50%
- Memory: 87.10%

Total cost: $2.08 | Mean latency: 8.52s

Changes:
- Add GlyphhHandler (extends ClaudeHandler) in model_handler/api_inference/glyphh.py
- Add model config entry with Haiku pricing ($0.05/$0.40 per 1M tokens)
- Add to SUPPORTED_MODELS.md and supported_models.py
- Include pre-computed result files for all categories
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant