README.md

metadata

license: apache-2.0
base_model: ibm-granite/granite-4.0-h-micro
tags:
  - microtensor
  - ctx_compress
extra_gated_heading: Request access to this model
extra_gated_prompt: >-
  This model was ranked in the Agent session compaction arena on Microtensor.
  Its files open once Microtensor approves your request. Tell us who you are and
  what you will use it for.
extra_gated_fields:
  Name: text
  Organisation: text
  Intended use: text
  I will use this model only under its licence: checkbox
extra_gated_button_content: Request access

Agent session compaction, epoch 0

This model ranked in the top 3 of a pool in the Agent session compaction arena (ctx_compress) on Microtensor (testnet), epoch 0. The weights, tokenizer and config here are byte for byte the files that were scored.

Placement	Score
1st in the Model pool	0.871

Results

Pool	Score	Lower bound	Items	Eligible
Model	0.871	0.867	25	yes
System	0.000	0.785	25	no

Spot check

The coordinator reran 40 sampled model calls on the profile card and 40 matched token for token. First token p95 5,457 ms, per token p95 8.3 ms, within the speed limits: yes, within the memory limit: yes.

Use it

from transformers import AutoModelForCausalLM, AutoTokenizer

tokenizer = AutoTokenizer.from_pretrained("microtensor-io/ctx_compress-e7528c49")
model = AutoModelForCausalLM.from_pretrained("microtensor-io/ctx_compress-e7528c49")

What is in this repo

Path	Contents
root	the scored model files, unchanged
README.miner.md	the miner's own README, unchanged
system/	the rest of the scored system: manifest, harness and escalation code
training/	the miner's training record, when the miner published one
certificates/	one signed certificate per epoch this artifact placed in
evidence/	the arena definition and links to the public records

Verify it

1. Take the SHA-256 of each weight file (.safetensors or .gguf), write one line per file as the file name, a space and the hash, sort the lines by file name, join them with newlines and take the SHA-256 of that text. It must equal e7528c4970d1dd273da438528078d8f4c86d3b6651882e5183ccbbfee5b08b7f.
2. Open certificates/e0.json, remove the coordinator_sig field, write the rest as JSON with sorted keys and no spaces, and take its SHA-256 as hex. The signature is over the text microtensor-certificate: followed by that hash, made with the sr25519 key of the coordinator hotkey 5D4cKxE8nX4jqJ7aMrrEhkXf7r6chouZbbn4ivQa9emk84m1.
3. Compare the scores with the public records listed in evidence/links.json.

Certificate hash: b6fb111bb86c0f48da14ad9468d581a9008f7a9ebdc1e501e682d3df85cc8975

The certificate covers the measured results in this arena and epoch only.

Source model: Samonte/mt_slm_ctx_compress@698535aacbd7d3555979fc0b7613d6b62dcfc895 by Samonte.